tooluniverse-target-research

name: tooluniverse-target-research description: Gather comprehensive biological target intelligence from 9 parallel research paths covering protein info, structure, interactions, pathways, expression, variants, drug interactions, and literature. Features collision-aware searches, evidence grading (T1-T4), explicit Open Targets coverage, and mandatory completeness auditing. Use when users ask about drug targets, proteins, genes, or need target validation, druggability assessment, or comprehensive target profiling.

Comprehensive Target Intelligence Gatherer

Gather complete target intelligence by exploring 9 parallel research paths. Supports targets identified by gene symbol, UniProt accession, Ensembl ID, or gene name.

KEY PRINCIPLES:

1. Report-first approach - Create report file FIRST, then populate progressively
2. Tool parameter verification - Verify params via get_tool_info before calling unfamiliar tools
3. Evidence grading - Grade all claims by evidence strength (T1-T4)
4. Citation requirements - Every fact must have inline source attribution
5. Mandatory completeness - All sections must exist with data minimums or explicit "No data" notes
6. Disambiguation first - Resolve all identifiers before research
7. Negative results documented - "No drugs found" is data; empty sections are failures
8. Collision-aware literature search - Detect and filter naming collisions
9. English-first queries - Always use English terms in tool calls, even if the user writes in another language. Translate gene names, disease names, and search terms to English. Only try original-language terms as a fallback if English returns no results. Respond in the user's language

Phase 0: Tool Parameter Verification (CRITICAL)

BEFORE calling ANY tool for the first time, verify its parameters:

# Always check tool params to prevent silent failures
tool_info = tu.tools.get_tool_info(tool_name="Reactome_map_uniprot_to_pathways")
# Reveals: takes `id` not `uniprot_id`

Known Parameter Corrections (Updated)

GTEx Versioned ID Fallback (CRITICAL)

GTEx often requires versioned Ensembl IDs. If ENSG00000123456 returns empty:

# Step 1: Get gene info with version
gene_info = tu.tools.ensembl_lookup_gene(id=ensembl_id, species="human")
version = gene_info.get('version', 1)

# Step 2: Try versioned ID
versioned_id = f"{ensembl_id}.{version}"  # e.g., "ENSG00000123456.12"
result = tu.tools.GTEx_get_median_gene_expression(
    gencode_id=versioned_id,
    operation="median"
)

When to Use This Skill

Apply when users:

• Ask about a drug target, protein, or gene
• Need target validation or assessment
• Request druggability analysis
• Want comprehensive target profiling
• Ask "what do we know about [target]?"
• Need target-disease associations
• Request safety profile for a target

Critical Workflow Requirements

1. Report-First Approach (MANDATORY)

DO NOT show the search process or tool outputs to the user. Instead:

1. Create the report file FIRST - Before any data collection:

   • File name: [TARGET]_target_report.md
   • Initialize with all 14 section headers
   • Add placeholder: [Researching...] in each section

2. Progressively update the report - As you gather data:

   • Update each section immediately after retrieving data
   • Replace [Researching...] with actual content
   • Include "No data returned" when tools return empty results

3. Methodology in appendix only - If user requests methodology details, create separate [TARGET]_methods_appendix.md

2. Evidence Grading System (MANDATORY)

CRITICAL: Grade every claim by evidence strength.

Evidence Tiers

Required Evidence Grading Locations

Evidence grades MUST appear in:

1. Executive Summary - Key disease claims graded
2. Section 8.2 Disease Associations - Every disease link graded with source type
3. Section 11 Literature - Key papers table with evidence tier
4. Section 13 Recommendations - Scorecard items reference evidence quality

Per-Section Evidence Summary

---
**Evidence Quality for this Section**: Strong
- Mechanistic (T1): 12 papers
- Functional (T2): 8 papers
- Association (T3): 15 papers
- Mention (T4): 23 papers
**Data Gaps**: No CRISPR data; mouse KO phenotypes limited
---

3. Citation Requirements (MANDATORY)

Every piece of information MUST include its source:

EGFR mutations cause lung adenocarcinoma [★★★: PMID:15118125, activating mutations 
in patients]. *Source: ClinVar, CIViC*

Core Strategy: 9 Research Paths

Execute 9 research paths (Path 0 is always first):

Target Query (e.g., "EGFR" or "P00533")
│
├─ IDENTIFIER RESOLUTION (always first)
│   └─ Check if GPCR → GPCRdb_get_protein
│
├─ PATH 0: Open Targets Foundation (ALWAYS FIRST - fills gaps in all other paths)
│
├─ PATH 1: Core Identity (names, IDs, sequence, organism)
│   └─ InterProScan_scan_sequence for novel domain prediction (NEW)
├─ PATH 2: Structure & Domains (3D structure, domains, binding sites)
│   └─ If GPCR: GPCRdb_get_structures (active/inactive states)
├─ PATH 3: Function & Pathways (GO terms, pathways, biological role)
├─ PATH 4: Protein Interactions (PPI network, complexes)
├─ PATH 5: Expression Profile (tissue expression, single-cell)
├─ PATH 6: Variants & Disease (mutations, clinical significance)
│   └─ DisGeNET_search_gene for curated gene-disease associations
├─ PATH 7: Drug Interactions (known drugs, druggability, safety)
│   ├─ Pharos_get_target for TDL classification (Tclin/Tchem/Tbio/Tdark)
│   ├─ BindingDB_get_ligands_by_uniprot for known ligands (NEW)
│   ├─ PubChem_search_assays_by_target_gene for HTS data (NEW)
│   ├─ If GPCR: GPCRdb_get_ligands (curated agonists/antagonists)
│   └─ DepMap_get_gene_dependencies for target essentiality
└─ PATH 8: Literature & Research (publications, trends)

Identifier Resolution (Phase 1)

CRITICAL: Resolve ALL identifiers before any research path.

def resolve_target_ids(tu, query):
    """
    Resolve target query to ALL needed identifiers.
    Returns dict with: query, uniprot, ensembl, ensembl_version, symbol, 
    entrez, chembl_target, hgnc
    """
    ids = {
        'query': query, 
        'uniprot': None, 
        'ensembl': None, 
        'ensembl_versioned': None,  # For GTEx
        'symbol': None,
        'entrez': None,
        'chembl_target': None,
        'hgnc': None,
        'full_name': None,
        'synonyms': []
    }
    
    # [Resolution logic based on input type]
    # ... (see current implementation)
    
    # CRITICAL: Get versioned Ensembl ID for GTEx
    if ids['ensembl']:
        gene_info = tu.tools.ensembl_lookup_gene(id=ids['ensembl'], species="human")
        if gene_info and gene_info.get('version'):
            ids['ensembl_versioned'] = f"{ids['ensembl']}.{gene_info['version']}"
        
        # Also get synonyms for literature collision detection
        ids['full_name'] = gene_info.get('description', '').split(' [')[0]
    
    # Get UniProt alternative names for synonyms
    if ids['uniprot']:
        alt_names = tu.tools.UniProt_get_alternative_names_by_accession(accession=ids['uniprot'])
        if alt_names:
            ids['synonyms'].extend(alt_names)
    
    return ids

GPCR Target Detection (NEW)

~35% of approved drugs target GPCRs. After identifier resolution, check if target is a GPCR:

def check_gpcr_target(tu, ids):
    """
    Check if target is a GPCR and retrieve specialized data.
    Call after identifier resolution.
    """
    symbol = ids.get('symbol', '')
    
    # Build GPCRdb entry name
    entry_name = f"{symbol.lower()}_human"
    
    gpcr_info = tu.tools.GPCRdb_get_protein(
        operation="get_protein",
        protein=entry_name
    )
    
    if gpcr_info.get('status') == 'success':
        # Target is a GPCR - get specialized data
        
        # Get structures with receptor state
        structures = tu.tools.GPCRdb_get_structures(
            operation="get_structures",
            protein=entry_name
        )
        
        # Get known ligands (critical for binder projects)
        ligands = tu.tools.GPCRdb_get_ligands(
            operation="get_ligands",
            protein=entry_name
        )
        
        # Get mutation data
        mutations = tu.tools.GPCRdb_get_mutations(
            operation="get_mutations",
            protein=entry_name
        )
        
        return {
            'is_gpcr': True,
            'gpcr_family': gpcr_info['data'].get('family'),
            'gpcr_class': gpcr_info['data'].get('receptor_class'),
            'structures': structures.get('data', {}).get('structures', []),
            'ligands': ligands.get('data', {}).get('ligands', []),
            'mutations': mutations.get('data', {}).get('mutations', []),
            'ballesteros_numbering': True  # GPCRdb provides this
        }
    
    return {'is_gpcr': False}

GPCRdb Report Section (add to Section 2 for GPCR targets):

### 2.x GPCR-Specific Data (GPCRdb)

**Receptor Class**: Class A (Rhodopsin-like)  
**GPCR Family**: Adrenoceptors  

**Structures by State**:
| PDB ID | State | Resolution | Ligand | Year |
|--------|-------|------------|--------|------|
| 3SN6 | Active | 3.2Å | Agonist (BI-167107) | 2011 |
| 2RH1 | Inactive | 2.4Å | Antagonist (carazolol) | 2007 |

**Known Ligands**: 45 agonists, 32 antagonists, 8 allosteric modulators  
**Key Binding Site Residues** (Ballesteros-Weinstein): 3.32, 5.42, 6.48, 7.39

Collision Detection for Literature Search

Before literature search, detect naming collisions:

def detect_collisions(tu, symbol, full_name):
    """
    Detect if gene symbol has naming collisions in literature.
    Returns negative filter terms if collisions found.
    """
    # Search by symbol in title
    results = tu.tools.PubMed_search_articles(
        query=f'"{symbol}"[Title]',
        limit=20
    )
    
    # Check if >20% are off-topic
    off_topic_terms = []
    for paper in results.get('articles', []):
        title = paper.get('title', '').lower()
        # Check if title mentions biology/protein/gene context
        bio_terms = ['protein', 'gene', 'cell', 'expression', 'mutation', 'kinase', 'receptor']
        if not any(term in title for term in bio_terms):
            # Extract potential collision terms
            # e.g., "JAK" might collide with "Just Another Kinase" jokes
            # e.g., "WDR7" might collide with other WDR family members in certain contexts
            pass
    
    # Build negative filter
    collision_filter = ""
    if off_topic_terms:
        collision_filter = " NOT " + " NOT ".join(off_topic_terms)
    
    return collision_filter

PATH 0: Open Targets Foundation (ALWAYS FIRST)

Objective: Populate baseline data for Sections 5, 8, 9, 10, 11 before specialized queries.

CRITICAL: Open Targets provides the most comprehensive aggregated data. Query ALL these endpoints:

Path 0 Implementation

def path_0_open_targets(tu, ids):
    """
    Open Targets foundation data - fills gaps for sections 5, 6, 8, 9, 10, 11.
    ALWAYS run this first.
    """
    ensembl_id = ids['ensembl']
    if not ensembl_id:
        return {'status': 'skipped', 'reason': 'No Ensembl ID'}
    
    results = {}
    
    # 1. Diseases & Phenotypes (Section 8)
    diseases = tu.tools.OpenTargets_get_diseases_phenotypes_by_target_ensemblId(
        ensemblId=ensembl_id
    )
    results['diseases'] = diseases if diseases else {'note': 'No disease associations returned'}
    
    # 2. Tractability (Section 9)
    tractability = tu.tools.OpenTargets_get_target_tractability_by_ensemblId(
        ensemblId=ensembl_id
    )
    results['tractability'] = tractability if tractability else {'note': 'No tractability data returned'}
    
    # 3. Safety Profile (Section 10)
    safety = tu.tools.OpenTargets_get_target_safety_profile_by_ensemblId(
        ensemblId=ensembl_id
    )
    results['safety'] = safety if safety else {'note': 'No safety liabilities identified'}
    
    # 4. Interactions (Section 6)
    interactions = tu.tools.OpenTargets_get_target_interactions_by_ensemblId(
        ensemblId=ensembl_id
    )
    results['interactions'] = interactions if interactions else {'note': 'No interactions returned'}
    
    # 5. GO Annotations (Section 5)
    go_terms = tu.tools.OpenTargets_get_target_gene_ontology_by_ensemblId(
        ensemblId=ensembl_id
    )
    results['go_terms'] = go_terms if go_terms else {'note': 'No GO annotations returned'}
    
    # 6. Publications (Section 11)
    publications = tu.tools.OpenTargets_get_publications_by_target_ensemblId(
        ensemblId=ensembl_id
    )
    results['publications'] = publications if publications else {'note': 'No publications returned'}
    
    # 7. Mouse Models (Section 8/10)
    mouse_models = tu.tools.OpenTargets_get_biological_mouse_models_by_ensemblId(
        ensemblId=ensembl_id
    )
    results['mouse_models'] = mouse_models if mouse_models else {'note': 'No mouse model data returned'}
    
    # 8. Chemical Probes (Section 9)
    probes = tu.tools.OpenTargets_get_chemical_probes_by_target_ensemblId(
        ensemblId=ensembl_id
    )
    results['chemical_probes'] = probes if probes else {'note': 'No chemical probes available'}
    
    # 9. Associated Drugs (Section 9)
    drugs = tu.tools.OpenTargets_get_associated_drugs_by_target_ensemblId(
        ensemblId=ensembl_id
    )
    results['drugs'] = drugs if drugs else {'note': 'No approved/trial drugs found'}
    
    return results

Negative Results Are Data

CRITICAL: Always document when a query returns empty:

### 9.3 Chemical Probes

**Status**: No validated chemical probes available for this target.
*Source: OpenTargets_get_chemical_probes_by_target_ensemblId returned empty*

**Implication**: Tool compound development would be needed for chemical biology studies.

PATH 2: Structure & Domains (Enhanced)

Objective: Robust structure coverage using 3-step chain.

3-Step Structure Search Chain

Do NOT rely solely on PDB text search. Use this chain:

def path_structure_robust(tu, ids):
    """
    Robust structure search using 3-step chain.
    """
    structures = {'pdb': [], 'alphafold': None, 'domains': [], 'method_notes': []}
    
    # STEP 1: UniProt PDB Cross-References (most reliable)
    if ids['uniprot']:
        entry = tu.tools.UniProt_get_entry_by_accession(accession=ids['uniprot'])
        pdb_xrefs = [x for x in entry.get('uniProtKBCrossReferences', []) 
                    if x.get('database') == 'PDB']
        for xref in pdb_xrefs:
            pdb_id = xref.get('id')
            # Get details for each PDB
            pdb_info = tu.tools.get_protein_metadata_by_pdb_id(pdb_id=pdb_id)
            if pdb_info:
                structures['pdb'].append(pdb_info)
        structures['method_notes'].append(f"Step 1: {len(pdb_xrefs)} PDB cross-refs from UniProt")
    
    # STEP 2: Sequence-based PDB Search (catches missing annotations)
    if ids['uniprot'] and len(structures['pdb']) < 5:
        sequence = tu.tools.UniProt_get_sequence_by_accession(accession=ids['uniprot'])
        if sequence and len(sequence) < 1000:  # Reasonable length for search
            similar = tu.tools.PDB_search_similar_structures(
                sequence=sequence[:500],  # Use first 500 AA if long
                identity_cutoff=0.7
            )
            if similar:
                for hit in similar[:10]:  # Top 10 similar
                    if hit['pdb_id'] not in [s.get('pdb_id') for s in structures['pdb']]:
                        structures['pdb'].append(hit)
        structures['method_notes'].append(f"Step 2: Sequence search (identity ≥70%)")
    
    # STEP 3: Domain-based Search (for multi-domain proteins)
    if ids['uniprot']:
        domains = tu.tools.InterPro_get_protein_domains(uniprot_accession=ids['uniprot'])
        structures['domains'] = domains if domains else []
        
        # For large proteins with domains, search by domain sequence windows
        if len(structures['pdb']) < 3 and domains:
            for domain in domains[:3]:  # Top 3 domains
                domain_name = domain.get('name', '')
                # Could search PDB by domain name
                domain_hits = tu.tools.PDB_search_by_keyword(query=domain_name, limit=5)
                if domain_hits:
                    structures['method_notes'].append(f"Step 3: Domain '{domain_name}' search")
    
    # AlphaFold (always check)
    alphafold = tu.tools.alphafold_get_prediction(uniprot_accession=ids['uniprot'])
    structures['alphafold'] = alphafold if alphafold else {'note': 'No AlphaFold prediction'}
    
    # IMPORTANT: Document limitations
    if not structures['pdb']:
        structures['limitation'] = "No direct PDB hit does NOT mean no structure exists. Check: (1) structures under different UniProt entries, (2) homolog structures, (3) domain-only structures."
    
    return structures

Structure Section Output Format

### 4.1 Experimental Structures (PDB)

**Total PDB Entries**: 23 structures *(Source: UniProt cross-references)*
**Search Method**: 3-step chain (UniProt xrefs → sequence search → domain search)

| PDB ID | Resolution | Method | Ligand | Coverage | Year |
|--------|------------|--------|--------|----------|------|
| 1M17 | 2.6Å | X-ray | Erlotinib | 672-998 | 2002 |
| 3POZ | 2.8Å | X-ray | Gefitinib | 696-1022 | 2010 |

**Note**: "No direct PDB hit" ≠ "no structure exists". Check homologs and domain structures.

PATH 5: Expression Profile (Enhanced)

GTEx with Versioned ID Fallback

def path_expression(tu, ids):
    """
    Expression data with GTEx versioned ID fallback.
    """
    results = {'gtex': None, 'hpa': None, 'failed_tools': []}
    
    # GTEx with fallback
    ensembl_id = ids['ensembl']
    versioned_id = ids.get('ensembl_versioned')
    
    # Try unversioned first
    gtex_result = tu.tools.GTEx_get_median_gene_expression(
        gencode_id=ensembl_id,
        operation="median"
    )
    
    # Fallback to versioned if empty
    if not gtex_result or gtex_result.get('data') == []:
        if versioned_id:
            gtex_result = tu.tools.GTEx_get_median_gene_expression(
                gencode_id=versioned_id,
                operation="median"
            )
            if gtex_result and gtex_result.get('data'):
                results['gtex'] = gtex_result
                results['gtex_note'] = f"Used versioned ID: {versioned_id}"
        
        if not results.get('gtex'):
            results['failed_tools'].append({
                'tool': 'GTEx_get_median_gene_expression',
                'tried': [ensembl_id, versioned_id],
                'fallback': 'See HPA data below'
            })
    else:
        results['gtex'] = gtex_result
    
    # HPA (always query as backup)
    hpa_result = tu.tools.HPA_get_rna_expression_by_source(ensembl_id=ensembl_id)
    results['hpa'] = hpa_result if hpa_result else {'note': 'No HPA RNA data'}
    
    return results

Human Protein Atlas - Extended Expression (NEW)

HPA provides comprehensive protein expression data including tissue-level, cell-level, and cell line expression.

def get_hpa_comprehensive_expression(tu, gene_symbol):
    """
    Get comprehensive expression data from Human Protein Atlas.
    
    Provides:
    - Tissue expression (protein and RNA)
    - Subcellular localization
    - Cell line expression comparison
    - Tissue specificity
    """
    
    # 1. Search for gene to get IDs
    gene_info = tu.tools.HPA_search_genes_by_query(search_query=gene_symbol)
    
    if not gene_info:
        return {'error': f'Gene {gene_symbol} not found in HPA'}
    
    # 2. Get tissue expression with specificity
    tissue_search = tu.tools.HPA_generic_search(
        search_query=gene_symbol,
        columns="g,gs,rnat,rnatsm,scml,scal",  # Gene, synonyms, tissue specificity, subcellular
        format="json"
    )
    
    # 3. Compare expression in cancer cell lines vs normal tissue
    cell_lines = ['a549', 'mcf7', 'hela', 'hepg2', 'pc3']
    cell_line_expression = {}
    
    for cell_line in cell_lines:
        try:
            expr = tu.tools.HPA_get_comparative_expression_by_gene_and_cellline(
                gene_name=gene_symbol,
                cell_line=cell_line
            )
            cell_line_expression[cell_line] = expr
        except:
            continue
    
    return {
        'gene_info': gene_info,
        'tissue_data': tissue_search,
        'cell_line_expression': cell_line_expression,
        'source': 'Human Protein Atlas'
    }

HPA Expression Output for Report:

### Tissue Expression Profile (Human Protein Atlas)

| Tissue | Protein Level | RNA nTPM | Specificity |
|--------|---------------|----------|-------------|
| Brain | High | 45.2 | Enriched |
| Liver | Medium | 23.1 | Enhanced |
| Kidney | Low | 8.4 | Not detected |

**Subcellular Localization**: Cytoplasm, Plasma membrane

### Cancer Cell Line Expression

| Cell Line | Cancer Type | Expression | vs Normal |
|-----------|-------------|------------|-----------|
| A549 | Lung | High | Elevated |
| MCF7 | Breast | Medium | Similar |
| HeLa | Cervical | High | Elevated |

*Source: Human Protein Atlas via `HPA_search_genes_by_query`, `HPA_get_comparative_expression_by_gene_and_cellline`*

Why HPA for Target Research:

• Drug target validation - Confirm expression in target tissue
• Safety assessment - Expression in essential organs
• Biomarker potential - Tissue-specific expression
• Cell line selection - Choose appropriate models

PATH 6: Variants & Disease (Enhanced)

6.1 ClinVar SNV vs CNV Separation

### 8.3 Clinical Variants (ClinVar)

#### Single Nucleotide Variants (SNVs)
| Variant | Clinical Significance | Condition | Review Status | PMID |
|---------|----------------------|-----------|---------------|------|
| p.L858R | Pathogenic | Lung cancer | 4 stars | 15118125 |
| p.T790M | Pathogenic | Drug resistance | 4 stars | 15737014 |

**Total Pathogenic SNVs**: 47

#### Copy Number Variants (CNVs) - Reported Separately
| Type | Region | Clinical Significance | Frequency |
|------|--------|----------------------|-----------|
| Amplification | 7p11.2 | Pathogenic | Common in cancer |

*Note: CNV data separated as it represents different mutation mechanism*

6.2 DisGeNET Integration (NEW)

DisGeNET provides curated gene-disease associations with evidence scores. Requires: DISGENET_API_KEY

def get_disgenet_associations(tu, ids):
    """
    Get gene-disease associations from DisGeNET.
    Complements Open Targets with curated association scores.
    """
    symbol = ids.get('symbol')
    if not symbol:
        return {'status': 'skipped', 'reason': 'No gene symbol'}
    
    # Get all disease associations for gene
    gda = tu.tools.DisGeNET_search_gene(
        operation="search_gene",
        gene=symbol,
        limit=50
    )
    
    if gda.get('status') != 'success':
        return {'status': 'error', 'message': 'DisGeNET query failed'}
    
    associations = gda.get('data', {}).get('associations', [])
    
    # Categorize by evidence strength
    strong = []     # score >= 0.7
    moderate = []   # score 0.4-0.7  
    weak = []       # score < 0.4
    
    for assoc in associations:
        score = assoc.get('score', 0)
        disease_name = assoc.get('disease_name', '')
        umls_cui = assoc.get('disease_id', '')
        
        entry = {
            'disease': disease_name,
            'umls_cui': umls_cui,
            'score': score,
            'evidence_index': assoc.get('ei'),
            'dsi': assoc.get('dsi'),  # Disease Specificity Index
            'dpi': assoc.get('dpi')   # Disease Pleiotropy Index
        }
        
        if score >= 0.7:
            strong.append(entry)
        elif score >= 0.4:
            moderate.append(entry)
        else:
            weak.append(entry)
    
    return {
        'total_associations': len(associations),
        'strong_associations': strong,
        'moderate_associations': moderate,
        'weak_associations': weak[:10],  # Limit weak
        'disease_pleiotropy': len(associations)  # How many diseases linked
    }

DisGeNET Report Section (add to Section 8 - Disease Associations):

### 8.x DisGeNET Gene-Disease Associations (NEW)

**Total Diseases Associated**: 47  
**Disease Pleiotropy Index**: High (gene linked to many disease types)

#### Strong Associations (Score ≥0.7)
| Disease | UMLS CUI | Score | Evidence Index |
|---------|----------|-------|----------------|
| Non-small cell lung cancer | C0007131 | 0.85 | 0.92 |
| Glioblastoma | C0017636 | 0.78 | 0.88 |

#### Moderate Associations (Score 0.4-0.7)
| Disease | UMLS CUI | Score | DSI |
|---------|----------|-------|-----|
| Breast cancer | C0006142 | 0.62 | 0.45 |

*Note: DisGeNET score integrates curated databases, GWAS, animal models, and literature*

Evidence Tier Assignment:

• DisGeNET Score ≥0.7 → Consider T2 evidence (multiple validated sources)
• DisGeNET Score 0.4-0.7 → Consider T3 evidence
• DisGeNET Score <0.4 → T4 evidence only

Extended Reference: For detailed tool tables, examples, and templates, read REFERENCE.md in this skill directory. The agent can access it via: read skills/tooluniverse-target-research/REFERENCE.md