ToolUniverse General Usage Strategies

Core Philosophy

1. Search widely - Don't assume you know all relevant tools; always search for more
2. Query multiple databases - Cross-reference data across sources
3. Multi-hop persistence - Many answers require 3-5 tool calls in sequence
4. Never give up easily - If one tool fails, try alternatives
5. Comprehensive reports - Use all available data; detail is valuable
6. English-first queries - Always use English terms in tool calls, even if the user writes in another language

Step 0: Clarify the Request Before Researching

When to Ask Clarifying Questions

When NOT to Ask

• "What is the structure of EGFR kinase domain?" - Clear entity + clear data type
• "Find all drugs targeting BRAF V600E" - Specific variant + clear task
• "Research Alzheimer's disease comprehensively" - Broad but unambiguous

Clarification Checklist

1. Entity - Exactly which gene/protein/drug/disease?
2. Species - Human unless stated otherwise
3. Scope - Comprehensive profile or specific aspect?
4. Output - Report, data table, quick answer, or comparison?

Strategy 1: Exhaustive Tool Discovery

Tool Discovery Methods

Tool_Finder_Keyword

Tool_Finder_LLM

Tool_Finder

Discovery Best Practices

Minimum Discovery Queries

1. Main topic query:

   [your topic]

2. Related terms:

   [synonym1]

   ,

   [synonym2]

3. Database-specific:

   [relevant database names]

4. Data type specific:

   [required data types]

• "protein information"
• "gene expression"
• "drug target"
• "UniProt", "OpenTargets"
• "protein interaction"
• "variant mutation"

Strategy 2: Multi-Hop Tool Chains

Why Multi-Hop Matters

Common Multi-Hop Patterns

Pattern A: ID Resolution Chain

Name → ID → Data → Related Data

Example: Gene name to complete profile
1. gene_name → Ensembl ID
2. Ensembl ID → UniProt accession  
3. UniProt accession → Protein entry
4. UniProt accession → Domains
5. UniProt accession → Structure

Pattern B: Cross-Database Enrichment

Primary Data → Cross-reference → Enriched View

Example: Drug compound enrichment
1. drug_name → PubChem CID
2. drug_name → ChEMBL ID
3. CID → properties
4. ChEMBL ID → bioactivity
5. ChEMBL ID → targets
6. SMILES → ADMET predictions

Pattern C: Network Expansion

Seed Entity → Connected Entities → Entity Details

Example: Target interaction network
1. gene → protein interactions
2. For each interactor → gene info
3. Interactor → disease associations

Pattern D: Literature + Data Integration

Database Annotations → Literature Evidence → Synthesis

Example: Disease mechanism research
1. disease → associated genes
2. disease → phenotypes
3. disease → drugs
4. disease → literature
5. key papers → citations

Multi-Hop Persistence Rules

1. Don't stop at first result - One tool gives partial data; keep going
2. Follow cross-references - Use IDs from one tool to query others
3. Chain until complete - 5-10 tool calls for comprehensive answers is normal
4. Track all IDs - Save every identifier for potential future use

Strategy 3: Query Multiple Databases for Same Data

Database Redundancy Principle

Merge Results Strategy

1. Collect all results - Don't stop at first success
2. Note data source - Track where each datum came from
3. Handle conflicts - Document when sources disagree
4. Prefer curated - Weight RefSeq over GenBank, UniProt over predictions

Strategy 3.1: Abstract Search vs Full-Text Search (Literature)

rs58542926

Quick rule

• If your keywords look like body-only terms (rsIDs, figure/table references, “Supplementary Table”), use full-text-aware tools first.

Tools that can match full text (indexed or retrieved)

PMC_search_papers

EuropePMC_search_articles

require_has_ft=true

fulltext_terms=[...]

HAS_FT:Y

BODY:\"...\"

EuropePMC_get_fulltext_snippets

retrieval_trace

CORE_search_papers

CORE_get_fulltext_snippets

openalex_search_works

openalex_literature_search

require_has_fulltext

fulltext_terms

Recommended flow for body-only keywords

1. Try

   PMC_search_papers

   and

   EuropePMC_search_articles

   (with

   require_has_ft

   +

   fulltext_terms

   ).
2. If you have a PMCID/PMID, use

   EuropePMC_get_fulltext_snippets

   to confirm the term is in the paper.
3. If you only have a PDF URL, use

   CORE_get_fulltext_snippets

   as a last resort, and treat HTTP

   200

   as “request succeeded”, not “PDF succeeded” (validate

   content_type

   ).

Strategy 4: Disambiguation First

Why Disambiguation Matters

Disambiguation Workflow

Step 1: Establish Canonical IDs
    gene_name → UniProt, Ensembl, NCBI Gene, ChEMBL target
    compound_name → PubChem CID, ChEMBL ID, SMILES
    disease_name → EFO ID, ICD-10, UMLS CUI

Step 2: Gather Synonyms
    All aliases, alternative names, historical names
    
Step 3: Detect Naming Collisions
    Search "[TERM]"[Title] → check if results are on-topic
    Build negative filters: NOT [collision_term]
    
Step 4: Species Confirmation
    Verify organism is correct (default: Homo sapiens)

ID Types by Entity

• Gene Symbol (EGFR, TP53)
• UniProt accession (P00533)
• Ensembl ID (ENSG00000146648)
• NCBI Gene ID (1956)
• ChEMBL Target ID (CHEMBL203)

• PubChem CID (2244)
• ChEMBL ID (CHEMBL25)
• SMILES string
• InChI/InChIKey

• EFO ID (EFO_0000249)
• ICD-10 code (G30)
• UMLS CUI (C0002395)
• SNOMED CT

Strategy 5: Never Give Up on Search

Failure Handling Protocol

Attempt 1: Primary tool
    ↓ fails
Wait briefly, then retry
    ↓ fails
Try fallback tool #1
    ↓ fails
Try fallback tool #2
    ↓ fails
Document as "unavailable" with reason

Common Fallback Chains

Alternative Search Strategies

• Try synonyms and aliases
• Use broader/narrower terms
• Try different databases

• Query related databases
• Use literature to find mentions
• Check if entity exists under different name

• Wait and retry
• Try same query on different database
• Use cached results if available

Strategy 6: Generate Comprehensive Reports

Report-First Approach

1. Create report structure FIRST - Define all sections before gathering data
2. Progressively update - Fill sections as data is gathered
3. Show findings, not process - Report results, not search methodology

Citation Requirements

## Protein Function

EGFR is a receptor tyrosine kinase that regulates cell growth.
*Source: UniProt (P00533)*

### Expression Profile
| Tissue | TPM | Source |
|--------|-----|--------|
| Skin | 156.3 | GTEx |
| Lung | 98.4 | GTEx |

Evidence Grading

ATP6V1A drives lysosomal acidification [★★★: PMID:12345678].
It has been implicated in cancer metabolism [★☆☆: TCGA data].

Mandatory Completeness

## Pathogen Involvement
No pathogen interactions identified in literature or databases.
*Source: Literature search, UniProt annotations*

Report Quality Metrics

Strategy 7: Use Specialized Skills for Specific Tasks

Task-Specific Skill Selection

tooluniverse-chemical-compound-retrieval

tooluniverse-expression-data-retrieval

tooluniverse-protein-structure-retrieval

tooluniverse-sequence-retrieval

tooluniverse-disease-research

tooluniverse-drug-research

tooluniverse-literature-deep-research

tooluniverse-target-research

tooluniverse-pharmacovigilance

tooluniverse-precision-oncology

tooluniverse-rare-disease-diagnosis

tooluniverse-variant-interpretation

tooluniverse-binder-discovery

drug-repurposing

tooluniverse-protein-therapeutic-design

tooluniverse-infectious-disease

setup-tooluniverse

tooluniverse-sdk

When to Use This General Skill

• Need general guidance on ToolUniverse usage
• Task doesn't fit a specialized skill
• Need to combine multiple specialized workflows
• Exploring what's possible with ToolUniverse
• Learning ToolUniverse best practices

Strategy 8: Parallel Execution for Speed

When to Parallelize

Parallel Research Paths Example

1. Identity - Names, IDs, sequence
2. Structure - 3D structure, domains
3. Function - GO terms, pathways
4. Interactions - PPI network
5. Expression - Tissue expression
6. Variants - Genetic variation
7. Drugs - Known drugs, druggability
8. Literature - Publications, trends

Strategy 9: Iterative Completeness Check

The Completeness Loop

Gather initial data
    ↓
Review what you have
    ↓
Ask: "What aspects are still missing?"
    ↓
Identify gaps
    ↓
Search for tools to fill gaps
    ↓
Gather additional data
    ↓
Repeat until comprehensive

Universal Completeness Questions

1. Identity: Do I have all relevant identifiers and names?
2. Core data: Do I have the fundamental information for this entity type?
3. Context: Do I have surrounding/related information?
4. Relationships: Do I know what this connects to?
5. Variations: Do I know about variants, forms, or subtypes?
6. Evidence: Do I have supporting data from multiple sources?
7. Literature: Do I have recent publications on this topic?
8. Gaps: Have I documented what's unavailable?

Gap-Filling Strategies

When to Stop

• All relevant aspects have been addressed (even if "not found")
• Multiple sources queried for key data
• Gaps are documented, not ignored
• No obvious missing pieces remain

Self-Review Questions

• Have I searched for ALL relevant tools?
• Have I queried multiple databases?
• Have I followed cross-references?
• Have I checked recent literature?
• Have I documented what's unavailable?
• Is there any obvious gap I haven't addressed?
• Would someone reading this ask "but what about X?"

Quick Reference: Tool Categories

Protein & Gene Tools

Structure Tools

Drug & Compound Tools

Disease & Phenotype Tools

Expression Tools

Variant Tools

Pathway Tools

Literature Tools

Clinical Tools

Troubleshooting Common Issues

"Tool not found"

• Search for similar tools using Tool_Finder
• Check spelling of tool name
• Try alternative tools for same data type

"Empty results"

• Check spelling of query terms
• Try synonyms/aliases
• Try alternative databases
• Verify IDs are correct type

"Conflicting data"

• Note all sources
• Prefer curated databases
• Document the conflict in report
• Use evidence grading

"Incomplete picture"

• Search for more tools
• Query additional databases
• Follow cross-references
• Expand via literature

Strategy 10: English-First Tool Queries

Language Handling Rules

1. Default to English - All tool calls must use English search terms, entity names, and parameters
2. Translate non-English input - If the user's question is in Chinese, Japanese, Korean, or any other language, translate the relevant scientific terms to English before making tool calls
3. Respond in the user's language - While tools must be queried in English, deliver the final report/answer in the user's original language
4. Fallback to original language - Only if an English search returns no results, retry with the original-language terms
5. Check tool descriptions - A few tools may explicitly document multi-language support; use the original language only when the tool description says so

Examples

User (Chinese): "研究EGFR靶点"
  → Tool calls: use "EGFR", "epidermal growth factor receptor" (English)
  → Report: deliver in Chinese

User (Japanese): "メトホルミンの安全性プロファイル"
  → Tool calls: use "metformin", "safety profile" (English)
  → Report: deliver in Japanese

User (Korean): "알츠하이머병 관련 유전자"
  → Tool calls: use "Alzheimer's disease", "associated genes" (English)
  → Report: deliver in Korean

Why This Matters

Summary: The ToolUniverse Mindset