Drug Repurposing with ToolUniverse

Systematically identify and evaluate drug repurposing candidates using multiple computational strategies.

IMPORTANT: Always use English terms in tool calls. Respond in the user's language.

Reasoning Before Searching

Start by asking: WHY might this drug work for a new disease? Three strategies:

• (a) Same target: The drug's primary target is also involved in the new disease. This is the strongest hypothesis — use OpenTargets to check if the target has genetic evidence in both diseases before any other search.
• (b) Off-target activity: The drug has secondary targets or off-target effects that are relevant to the new disease. Check ChEMBL bioactivity data for all known targets of the drug, not just its primary one.
• (c) Shared pathways: The original indication and new disease share molecular pathways, even if the target itself is not genetically linked. Use Reactome and STRING to compare pathway overlap between diseases.

Each strategy uses different tools and has different evidentiary weight. Identify which strategy applies FIRST, then choose the corresponding workflow below. Do not run all three strategies blindly — reason about which is most plausible given the drug's mechanism.

LOOK UP DON'T GUESS: Never assume a drug hits a target, never assume a target is disease-relevant, never assume pathway overlap. Verify each link with tool calls.

Core Strategies

1. Target-Based: Disease targets -> Find drugs that modulate those targets
2. Compound-Based: Approved drugs -> Find new disease indications
3. Disease-Driven: Disease -> Targets -> Match to existing drugs
4. Signature-Based: Disease gene-expression signature -> Find drugs whose perturbation signature reverses it (connectivity-map reasoning). Use when the disease has a well-characterized differential-expression signature (e.g. from GEO/TCGA) but no single validated target — this strategy needs no target hypothesis at all, only up/down gene lists.

Workflow Overview

Phase 1: Disease & Target Analysis
  Get disease info (OpenTargets), find associated targets, get target details

Phase 2: Drug Discovery
  Search DrugBank, DGIdb, ChEMBL for drugs targeting disease-associated genes
  Get drug details, indications, pharmacology

Phase 3: Safety & Feasibility Assessment
  FDA warnings, FAERS adverse events, drug interactions, ADMET predictions

Phase 4: Literature Evidence
  PubMed, Europe PMC, clinical trials for existing evidence

Phase 5: Scoring & Ranking
  Composite score: target association + safety + literature + drug properties

See: PROCEDURES.md for detailed step-by-step procedures and code patterns.

Quick Start

from tooluniverse import ToolUniverse
tu = ToolUniverse()
tu.load_tools()

# Step 1: Get disease targets
disease_info = tu.tools.OpenTargets_get_disease_id_description_by_name(diseaseName="rheumatoid arthritis")
# Response nests ID at data.search.hits[0].id
disease_id = disease_info['data']['search']['hits'][0]['id']
targets = tu.tools.OpenTargets_get_associated_targets_by_disease_efoId(efoId=disease_id, size=10)

# Step 2: Find drugs for each target
# Response nests targets at data.disease.associatedTargets.rows
rows = targets['data']['disease']['associatedTargets']['rows']
for target in rows[:5]:
    gene = target['target']['approvedSymbol']
    drugs = tu.tools.DGIdb_get_drug_gene_interactions(genes=[gene])

Key ToolUniverse Tools

Disease & Target:

• OpenTargets_get_disease_id_description_by_name - Disease lookup
• OpenTargets_get_associated_targets_by_disease_efoId - Disease targets
• UniProt_get_entry_by_accession - Protein details

Drug Discovery:

• drugbank_get_drug_name_and_description_by_target_name - Drugs by target. Param: query= (NOT target_name=)
• drugbank_get_drug_name_and_description_by_indication - Drugs by indication. Param: query= (NOT indication=)
• DGIdb_get_drug_gene_interactions - Drug-gene interactions. Response path: data.data.genes.nodes[0].interactions
• ChEMBL_search_drugs / ChEMBL_get_drug_mechanisms - Drug search and MOA

Patent/IP (USPTO Patent File Wrapper — see "Patent Landscape Check" below):

• get_patent_overview_by_text_query - Keyword/title patent search
• get_patent_application_metadata, get_patent_term_adjustment_data, get_patent_continuity_data, get_patent_foreign_priority_data, get_associated_documents_metadata - Per-application detail lookups by application number

Drug Information (ALL DrugBank tools use query= as the search parameter, plus case_sensitive=False, exact_match=False, limit=N):

• drugbank_get_drug_basic_info_by_drug_name_or_id - Basic info. Param: query="drug_name"
• drugbank_get_indications_by_drug_name_or_drugbank_id - Approved indications. Param: query="drug_name"
• drugbank_get_pharmacology_by_drug_name_or_drugbank_id - Pharmacology. Param: query="drug_name"
• drugbank_get_targets_by_drug_name_or_drugbank_id - Drug targets. Param: query="drug_name"

Safety:

• FDA_get_warnings_and_cautions_by_drug_name - FDA warnings
• FAERS_search_reports_by_drug_and_reaction - Adverse events. Param: medicinalproduct= (NOT drug_name=)
• FAERS_count_death_related_by_drug - Serious outcomes. Param: medicinalproduct= (NOT drug_name=)
• drugbank_get_drug_interactions_by_drug_name_or_id - Interactions

Property Prediction:

• ADMETAI_predict_physicochemical_properties / ADMETAI_predict_toxicity - ADMET and toxicity

Pathway & Network Analysis:

• ReactomeAnalysis_pathway_enrichment - Pathway enrichment. Param: identifiers="SOD1\nTARDBP\nFUS" (newline-separated string, NOT array)
• STRING_get_network - Protein interaction networks. Param: identifiers="SOD1\rTARDBP\rFUS" (CR-separated string), species=9606
• CTD_get_gene_diseases - Curated gene-disease associations. Param: input_terms="gene_symbol" (NOT gene_symbol=)

Literature & Clinical Trials:

• PubMed_search_articles / EuropePMC_search_articles - Literature search
• search_clinical_trials - ClinicalTrials.gov search. Use condition for disease name. The intervention filter is strict and may miss trials — use query_term for broader drug-name matching as fallback.

CNS diseases note: For neurological indications (ALS, Alzheimer's, Parkinson's), prioritize BBB-penetrant candidates. Use ChEMBL molecular properties (MW < 500, PSA < 90) as BBB proxy since ADMETAI_predict_BBB_penetrance may require the tooluniverse[ml] extra. Consider route of administration (oral preferred for patients with swallowing difficulty) and sex-specific effects from preclinical models.

Scoring & Decision Framework

Repurposing Viability Score (0-100)

Classification:

• 80-100: Strong candidate — proceed to clinical evaluation
• 60-79: Promising — worth preclinical validation or retrospective study
• 40-59: Speculative — needs significant additional evidence
• <40: Weak — likely not worth pursuing without new mechanistic insight

Evidence Grading for Repurposing

How to Interpret and Combine Results

After running Phases 1-4, synthesize by answering:

1. Is the target validated for this disease? Check OpenTargets association score (>0.5 = strong). Cross-reference with genetic evidence (GWAS hits, rare variant studies). If target association is only pathway-level, the repurposing hypothesis is speculative.

2. Does the drug actually hit the target at achievable doses? Check ChEMBL IC50/Ki values. If the drug's affinity for the new target is >10x weaker than for its original target, clinical efficacy is unlikely at safe doses.

3. What's the safety margin? Compare the dose needed for the new indication to the approved dose. If higher doses are needed, safety data from the original indication may not apply.

4. Is there prior clinical evidence? A Phase II trial for the new indication (even failed) is more informative than 100 computational predictions. Check search_clinical_trials first.

5. What's the competitive landscape? If better drugs already exist for the disease, repurposing offers little value. Check DrugBank indications for approved therapies.

Best Practices

1. Check clinical trials FIRST: search_clinical_trials(condition="[disease]", intervention="[drug]") — if a trial already exists, start there
2. Validate targets with genetics: Genetic evidence (GWAS, rare variants) is the strongest predictor of successful drug development
3. Safety first: Prioritize approved drugs with known safety profiles
4. Dose matters: A drug that hits a disease target at 100x its approved dose is not a repurposing candidate
5. Mechanism over correlation: Network proximity alone is insufficient — explain WHY the drug should work
6. Consider IP and formulation: Generic drugs are easier to repurpose but harder to fund trials for. See "Patent Landscape Check" below to verify IP status rather than assuming a drug is off-patent because it's old or generic in one market — patent term adjustments and continuation filings can extend protection well past the nominal 20-year term.

Patent Landscape Check (USPTO)

Before prioritizing a repurposing candidate, check whether the new use (or the compound itself) is already claimed by a live patent. Six tools wrap the USPTO Patent File Wrapper API for this:

Requires USPTO_API_KEY (free, register at https://data.uspto.gov/myodp). Without a valid key every call returns HTTP 403 — verified live: get_patent_overview_by_text_query and get_patent_application_metadata both return HTTP Error: 403 with USPTO's own guidance that a 403 means key rejection (invalid, or a newly issued key still activating), not a malformed query. A 404 on a specific patent number means that patent is outside the Patent File Wrapper dataset's coverage, not that the tool is broken.

Workflow: search by drug/compound name or title keyword with get_patent_overview_by_text_query (wrap multi-word phrases in escaped double quotes for exact matching) → take an applicationNumberText from a hit → pull term-adjustment and continuity data to see the actual remaining protection window, since continuations and PTA both push the real expiration later than the nominal grant-date + 20-years math suggests. A drug whose base composition patent looks expired may still be covered by a use-patent or formulation continuation — always check continuity data before concluding a compound is open for repurposing without licensing.

Signature-Based Repurposing (LINCS / L1000FWD)

For Strategy 4, reverse a disease's own gene-expression signature rather than reasoning through a single target:

Workflow: derive the disease's up/down gene sets first (e.g. from a published DE analysis, GEO dataset, or this repo's own tooluniverse-rnaseq-deseq2 output) → L1000FWD_sig_search(up_genes=disease_up, down_genes=disease_down, mode="reverse") → each hit returns combined_scores/pvals/qvals — rank by score, and treat a top reversal hit as an E4 (Computational) candidate needing the same target-validation and dose-feasibility scrutiny as any other strategy, not a shortcut past it. Cross-check any promising hit against LINCS_search_signatures(drug_name=<hit>) to see its signature directly rather than trusting the connectivity score alone.

Verified live: LINCS_search_signatures(drug_name="vorinostat") and L1000FWD_sig_search both return real, non-empty results — this is a working, queryable strategy, not a placeholder.

Cross-Knowledge-Graph Queries (NCATS Biomedical Data Translator)

For a fast, single-call sweep across ~15 knowledge providers at once (rather than querying OpenTargets/ChEMBL/Reactome one at a time), the NCATS Translator tools give a biolink-standardized alternative:

Verified live: resolve_entity(name="Alzheimer disease") → MONDO:0004975; query_associations(entity_id="MONDO:0004975", target_category="ChemicalEntity", predicate="treats") → real ranked hits (e.g. CHEBI:87631 "statin", score ~0.9). This is a good Strategy (c)-style breadth check — cheap to run early, but its ranked "score" is E4 (Computational) evidence like any other network-proximity signal, not a substitute for the target-genetics/ChEMBL-affinity/clinical-trial verification steps above.

Computational Procedure: Drug-Target Dose Feasibility Check

A drug that hits a new target only at 100x its approved dose is NOT a viable repurposing candidate. Use this procedure after identifying drug-target pairs:

# Drug-target dose feasibility analysis
# Uses ChEMBL bioactivity data from ToolUniverse
from tooluniverse import ToolUniverse

tu = ToolUniverse()
tu.load_tools()

def check_dose_feasibility(drug_name, original_target, new_target):
    """
    Compare drug's potency at original vs new target.
    If new_target IC50 > 10x original_target IC50, flag as unlikely feasible.
    """
    # Get bioactivity for original target
    orig = tu.run_one_function({
        'name': 'ChEMBL_search_activities',
        'arguments': {
            'molecule_chembl_id': drug_name,  # or search first
            'target_chembl_id': original_target,
            'limit': 10
        }
    })

    # Get bioactivity for new target
    new = tu.run_one_function({
        'name': 'ChEMBL_search_activities',
        'arguments': {
            'molecule_chembl_id': drug_name,
            'target_chembl_id': new_target,
            'limit': 10
        }
    })

    # Extract IC50/Ki values and compare
    # If new target requires >10x concentration → NOT FEASIBLE at safe doses
    # If new target is within 3x → PROMISING
    # If new target is within 1x → STRONG candidate
    pass  # Parse actual values from results

# Alternative: Quick Cmax check
# If published Cmax at approved dose < IC50 for new target → NOT FEASIBLE
# Cmax data can be found in:
#   - DrugBank pharmacology section
#   - DailyMed clinical pharmacology section
#   - PubMed PK studies

Key principle: The most common reason repurposing fails is insufficient drug exposure at the new target. Always check whether the drug's concentration at approved doses reaches the IC50 for the new target.

Troubleshooting

Reference Files

• REFERENCE.md - Detailed reference documentation
• EXAMPLES.md - Sample repurposing analyses
• PROCEDURES.md - Step-by-step procedures with code
• REPORT_TEMPLATE.md - Output report template
• Related skills: disease-intelligence-gatherer, chemical-compound-retrieval, tooluniverse-sdk