tooluniverse-small-molecule-discovery（工具宇宙小分子发现）

Small Molecule Discovery Skill

Systematic small molecule identification, characterization, and sourcing using PubChem, ChEMBL, BindingDB, ADMET-AI, SwissADME, eMolecules, and Enamine. Covers the full pipeline from compound name to structure, activity, ADMET properties, and commercial procurement.

Domain Reasoning

Drug-likeness is not a binary property. Lipinski's Rule of 5 was derived from orally administered, passively absorbed drugs and has many well-known exceptions: natural products, macrocycles, PROTACs, and many approved drugs violate one or more rules. The relevant question is not "does this pass Ro5?" but "does this compound's physicochemical profile match the requirements of the target, the intended route of administration, and the therapeutic context?" Focus on the specific requirements, not rigid rules.

LOOK UP DON'T GUESS

• Compound identity (CID, ChEMBL ID, SMILES): call PubChem_get_CID_by_compound_name and ChEMBL_search_molecules; do not assume IDs from memory.
• ADMET properties: run SwissADME_calculate_adme or ADMETAI_predict_* on the actual SMILES; do not estimate logP, TPSA, or bioavailability.
• Binding affinities against a target: query ChEMBL_search_activities or BindingDB_get_ligands_by_uniprot; never cite IC50 values from memory.
• Commercial availability: check eMolecules_search or Enamine_search_catalog; do not assume availability.

KEY PRINCIPLES:

1. Resolve identity first - Always get CID and ChEMBL ID before research
2. SMILES required for property prediction - Extract canonical SMILES from PubChem early
3. English names in tools - Use IUPAC or common English names; avoid abbreviations in tool calls
4. BindingDB is often unavailable - Fall back to ChEMBL activities when BindingDB times out
5. eMolecules/Enamine return URLs - These tools generate search URLs, not direct data; note this to user

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.

When to Use

• "Find information about compound X"
• "What is the drug-likeness of this SMILES?"
• "Show binding affinities for EGFR inhibitors"
• "Search for compounds similar to imatinib"
• "Is this compound commercially available?"
• "What are the ADMET properties of this molecule?"
• "Find ChEMBL activities for target Y"
• "Predict targets for this small molecule"

Key Tools

Workflow

Phase 1: Compound Identification

# Step 1: Name -> CID (PubChem canonical identity)
PubChem_get_CID_by_compound_name(compound_name="imatinib")
# -> CID: 5291

# Step 2: Get SMILES and properties (needed for all downstream tools)
PubChem_get_compound_properties_by_CID(
    cid="5291",
    properties="MolecularFormula,MolecularWeight,CanonicalSMILES,InChIKey,IUPACName"
)
# -> canonical SMILES, InChIKey (global identifier)

# Step 3: Get ChEMBL ID (for activity data)
ChEMBL_search_molecules(query="imatinib")
# -> ChEMBL ID (e.g., "CHEMBL941")

# Step 4: Get all synonyms (brand names, INN, etc.)
PubChem_get_compound_synonyms_by_CID(cid="5291")

ID resolution priority:

1. Start with PubChem CID (most universal)
2. Get ChEMBL ID (for bioactivity data)
3. Use canonical SMILES for structure-based searches and ADMET

Phase 2: Structure-Based Search

Similarity search (find analogs):

PubChem_search_compounds_by_similarity(
    smiles="CANONICAL_SMILES",
    threshold=85   # Tanimoto threshold 0-100; 85 = highly similar
)
# Returns: list of CIDs of similar compounds

ChEMBL_search_similar_molecules(query="CHEMBL941")  # Or SMILES
# Returns: ChEMBL entries sorted by similarity

Substructure search (find compounds containing a scaffold):

PubChem_search_compounds_by_substructure(smiles="SCAFFOLD_SMILES")
# Returns: CIDs of compounds containing the scaffold

Phase 3: Bioactivity and Binding Affinity

Get all activities for a compound (across all targets):

ChEMBL_search_activities(
    molecule_chembl_id="CHEMBL941",
    pchembl_value__gte=6,   # pIC50/Ki >= 6 = IC50/Ki <= 1 µM
    limit=50
)
# Returns: assay_type, target_name, pchembl_value, units

Get all ligands for a target:

# First find target ChEMBL ID
ChEMBL_search_targets(query="EGFR", organism="Homo sapiens")
# -> target_chembl_id, e.g., "CHEMBL203"

ChEMBL_get_target_activities(
    target_chembl_id="CHEMBL203"
)
# Returns: all compounds with binding data against this target

BindingDB (when available — often times out):

BindingDB_get_ligands_by_uniprot(uniprot_id="P00533")  # EGFR
# Returns: Ki, IC50, Kd data with literature references
# Note: BindingDB REST API is frequently unavailable; fall back to ChEMBL

pChEMBL Value interpretation:

Phase 4: Drug-likeness and ADMET

SwissADME (comprehensive, requires SMILES string — not list):

SwissADME_calculate_adme(
    operation="calculate_adme",
    smiles="CANONICAL_SMILES"
)
# Returns: physicochemical, lipophilicity, water solubility, pharmacokinetics,
#          drug-likeness scores (Lipinski, Veber, Egan, Muegge), PAINS alerts

SwissADME_check_druglikeness(
    operation="check_druglikeness",
    smiles="CANONICAL_SMILES"
)
# Returns: Lipinski/Veber/Egan pass/fail + lead-likeness

ADMET-AI (ML-based, requires SMILES as list — install tooluniverse[ml]):

ADMETAI_predict_physicochemical_properties(smiles=["CANONICAL_SMILES"])
ADMETAI_predict_bioavailability(smiles=["CANONICAL_SMILES"])
ADMETAI_predict_BBB_penetrance(smiles=["CANONICAL_SMILES"])
ADMETAI_predict_toxicity(smiles=["CANONICAL_SMILES"])
ADMETAI_predict_CYP_interactions(smiles=["CANONICAL_SMILES"])

Note: ADMET-AI requires pip install tooluniverse[ml]. If unavailable, use SwissADME as fallback.

Key drug-likeness rules:

• Lipinski Ro5: MW <= 500, logP <= 5, HBD <= 5, HBA <= 10 (oral drugs)
• Veber: TPSA <= 140 Ų, rotatable bonds <= 10 (oral bioavailability)
• Lead-like: MW <= 350, logP <= 3, HBD <= 3, HBA <= 6 (fragment/lead)

Phase 5: Target Prediction

When you have a novel compound and want to predict targets:

SwissTargetPrediction_predict(
    operation="predict",
    smiles="CANONICAL_SMILES"
)
# Returns: predicted protein targets with probability scores
# Note: SwissTargetPrediction uses structure-similarity to known drug-target pairs
# May time out for complex molecules

Phase 6: Commercial Availability

eMolecules (aggregates 200+ suppliers — returns search URL, not direct data):

eMolecules_search(query="compound_name")
# -> Returns search_url to visit on eMolecules.com

eMolecules_search_smiles(smiles="CANONICAL_SMILES")
# -> Returns URL for exact/similar structure search

Enamine (37B+ make-on-demand compounds — returns URL when API unavailable):

Enamine_search_catalog(query="compound_name")
# -> If API available: returns catalog entries with catalog_id, price
# -> If API unavailable: returns search_url for manual search

Enamine_search_smiles(smiles="CANONICAL_SMILES")
# -> Exact or similarity structure search

Enamine_get_libraries()
# -> Lists available Enamine screening collections

Note: eMolecules and Enamine APIs frequently return search URLs rather than live data. Present these to the user as "search here" links.

Tool Parameter Reference

Common Patterns

Pattern 1: Full Compound Profile

Input: Compound name (e.g., "imatinib")
Flow:
  1. PubChem_get_CID_by_compound_name -> CID + SMILES
  2. ChEMBL_search_molecules -> ChEMBL ID
  3. PubChem_get_compound_properties_by_CID -> physicochemical props
  4. SwissADME_calculate_adme / ADMETAI_predict_* -> ADMET profile
  5. ChEMBL_search_activities(molecule_chembl_id) -> binding data
  6. ChEMBL_get_drug_mechanisms -> MOA (if approved drug)
Output: Complete compound profile with identity, ADMET, and activity data

Pattern 2: Analog Discovery

Input: Reference compound SMILES
Flow:
  1. PubChem_search_compounds_by_similarity(smiles, threshold=85) -> similar CIDs
  2. ChEMBL_search_similar_molecules(query=smiles) -> ChEMBL analogs
  3. For each hit: PubChem_get_compound_properties_by_CID -> properties
  4. SwissADME_check_druglikeness -> filter by drug-likeness
Output: Ranked list of analogs with activity data and drug-likeness scores

Pattern 3: Target-Based Compound Search

Input: Target name (e.g., "EGFR")
Flow:
  1. ChEMBL_search_targets(query="EGFR", organism="Homo sapiens") -> target_chembl_id
  2. ChEMBL_get_target_activities(target_chembl_id) -> all ligands with Ki/IC50
  3. Filter by pchembl_value >= 7 (potent compounds)
  4. For top hits: SwissADME_check_druglikeness -> assess drug-likeness
  5. eMolecules_search(query=compound_name) -> check commercial availability
Output: Prioritized list of potent, drug-like, commercially available compounds

Pattern 4: ADMET Risk Assessment

Input: Novel compound SMILES
Flow:
  1. SwissADME_calculate_adme(operation="calculate_adme", smiles) -> full ADMET
  2. ADMETAI_predict_toxicity(smiles=[smiles]) -> hERG, DILI, mutagenicity
  3. ADMETAI_predict_CYP_interactions(smiles=[smiles]) -> drug-drug interaction risk
  4. ADMETAI_predict_BBB_penetrance(smiles=[smiles]) -> CNS penetration
Output: ADMET risk profile with flagged liabilities

Fallback Chains

Limitations

• BindingDB: REST API frequently times out; ChEMBL is the reliable alternative for binding data
• Enamine API: Returns HTTP 500 often; tool provides search URL as fallback
• eMolecules: No public API; tool generates search URLs only
• ADMET-AI: Requires pip install tooluniverse[ml]; not always available in base install
• SwissTargetPrediction: Web scraping-based; may time out for complex molecules
• SMILES format: ADMET-AI requires a list ["SMILES"]; SwissADME requires a string "SMILES"
• ChEMBL IDs: Always use full format "CHEMBL941", never just "941"