Chemical Safety & Toxicology Assessment

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 safety claims by evidence strength (T1-T4)
4. Citation requirements - Every toxicity finding must have inline source attribution
5. Mandatory completeness - All sections must exist with data minimums or explicit "No data" notes
6. Disambiguation first - Resolve compound identity (name -> SMILES, CID, ChEMBL ID) before analysis
7. Negative results documented - "No toxicity signals found" is data; empty sections are failures
8. Conservative risk assessment - When evidence is ambiguous, flag as "requires further investigation"
9. English-first queries - Always use English chemical/drug names in tool calls

Evidence Grading System (MANDATORY)

Grade every toxicity claim by evidence strength:

Required Evidence Grading Locations

Evidence grades MUST appear in:

1. Executive Summary - Key toxicity findings graded
2. Toxicity Predictions - Every ADMET-AI endpoint with confidence note
3. Regulatory Safety - FDA findings marked [T1]
4. Chemical-Gene Interactions - CTD data marked by curation status
5. Risk Assessment - Final risk classification with supporting evidence tiers

Core Strategy: 8 Research Dimensions

Chemical/Drug Query
|
+-- PHASE 0: Compound Disambiguation (ALWAYS FIRST)
|   +-- Resolve name -> SMILES, PubChem CID, ChEMBL ID
|   +-- Get molecular formula, weight, canonical structure
|
+-- PHASE 1: Predictive Toxicology (ADMET-AI)
|   +-- Mutagenicity (AMES)
|   +-- Hepatotoxicity (DILI, ClinTox)
|   +-- Carcinogenicity
|   +-- Acute toxicity (LD50)
|   +-- Skin reactions
|   +-- Stress response pathways
|   +-- Nuclear receptor activity
|
+-- PHASE 2: ADMET Properties
|   +-- Absorption: BBB penetrance, bioavailability
|   +-- Distribution: clearance, volume of distribution
|   +-- Metabolism: CYP interactions (1A2, 2C9, 2C19, 2D6, 3A4)
|   +-- Physicochemical: solubility, lipophilicity, pKa
|
+-- PHASE 3: Toxicogenomics (CTD)
|   +-- Chemical-gene interactions
|   +-- Chemical-disease associations
|   +-- Affected biological pathways
|
+-- PHASE 4: Regulatory Safety (FDA Labels)
|   +-- Boxed warnings (Black Box)
|   +-- Contraindications
|   +-- Adverse reactions
|   +-- Warnings and precautions
|   +-- Nonclinical toxicology
|
+-- PHASE 5: Drug Safety Profile (DrugBank)
|   +-- Toxicity data
|   +-- Contraindications
|   +-- Drug interactions affecting safety
|
+-- PHASE 6: Chemical-Protein Interactions (STITCH)
|   +-- Direct chemical-protein binding
|   +-- Interaction confidence scores
|   +-- Off-target effects
|
+-- PHASE 7: Structural Alerts (ChEMBL)
|   +-- Known toxic substructures (PAINS, Brenk)
|   +-- Structural alert flags
|
+-- SYNTHESIS: Integrated Risk Assessment
    +-- Aggregate all evidence tiers
    +-- Risk classification (Low/Medium/High/Critical)
    +-- Data gaps and recommendations

Phase 0: Compound Disambiguation (ALWAYS FIRST)

CRITICAL: Resolve compound identity before any analysis.

Input Types Handled

Resolution Steps

1. Input detection: Determine if input is name, SMILES, CID, or ChEMBL ID
   • SMILES: contains typical SMILES characters (=, #, [, ], (, ), c, n, o and no spaces in middle)
   • CID: numeric only
   • ChEMBL: starts with "CHEMBL"
   • Otherwise: treat as compound name
2. Name to CID: PubChem_get_CID_by_compound_name(name=<compound_name>)
3. CID to properties: PubChem_get_compound_properties_by_CID(cid=<cid>)
4. Extract SMILES: Get SMILES from PubChem properties (field: ConnectivitySMILES, CanonicalSMILES, or IsomericSMILES depending on response format)
5. Store resolved IDs: Maintain dict with name, smiles, cid, formula, weight, inchi

Disambiguation Output

## Compound Identity

| Property | Value |
|----------|-------|
| **Name** | Acetaminophen |
| **PubChem CID** | 1983 |
| **SMILES** | CC(=O)Nc1ccc(O)cc1 |
| **Formula** | C8H9NO2 |
| **Molecular Weight** | 151.16 |
| **InChI** | InChI=1S/C8H9NO2/... |

Phase 1: Predictive Toxicology (ADMET-AI)

When: SMILES is available (from Phase 0 or provided directly)

Objective: Run comprehensive AI-predicted toxicity endpoints

Tools Used

All ADMET-AI tools take the same parameter format:

Workflow

1. Call ADMETAI_predict_toxicity(smiles=[resolved_smiles])
2. Call ADMETAI_predict_stress_response(smiles=[resolved_smiles])
3. Call ADMETAI_predict_nuclear_receptor_activity(smiles=[resolved_smiles])
4. For each endpoint, interpret prediction:
   • Classification endpoints: Active (1) = toxic signal, Inactive (0) = no signal
   • Regression endpoints (LD50): Report numerical value with context
   • All predictions graded [T3] (computational prediction)

Decision Logic

• Multiple SMILES: Can batch up to ~10 SMILES in single call
• Failed prediction: If ADMET-AI fails, note "prediction unavailable" (don't fail entire report)
• Confidence: Note that AI predictions are [T3] evidence, not definitive
• hERG flag: If hERG = Active, flag prominently (cardiac safety risk)
• AMES flag: If AMES = Active, flag prominently (mutagenicity concern)
• DILI flag: If DILI = Active, flag prominently (liver toxicity concern)

Output Table

### Toxicity Predictions [T3]

| Endpoint | Prediction | Interpretation | Concern Level |
|----------|-----------|---------------|---------------|
| AMES Mutagenicity | Inactive | No mutagenic signal | Low |
| Carcinogenicity | Inactive | No carcinogenic signal | Low |
| ClinTox | Active | Clinical toxicity signal | HIGH |
| DILI | Active | Drug-induced liver injury risk | HIGH |
| LD50 (Zhu) | 2.45 log(mg/kg) | ~282 mg/kg (moderate) | Medium |
| Skin Reaction | Inactive | No skin sensitization signal | Low |
| hERG Inhibition | Active | Cardiac arrhythmia risk | HIGH |

*All predictions from ADMET-AI. Evidence tier: [T3] (computational prediction)*

Phase 2: ADMET Properties

When: SMILES is available

Objective: Full ADMET characterization beyond toxicity

Tools Used

Workflow

1. Call all 6 ADMET tools in parallel (independent calls)
2. Compile results into Absorption / Distribution / Metabolism / Excretion sections
3. Assess Lipinski Rule of 5 compliance from physicochemical properties
4. Flag drug-drug interaction risks from CYP inhibition profiles

Decision Logic

• BBB penetrant + toxicity: If BBB = Yes and any CNS toxicity endpoint active, flag as neurotoxicity risk
• Low bioavailability: If F20% = Low, note absorption concerns
• CYP inhibitor: If CYP3A4 inhibitor = Yes, flag high DDI risk
• Lipinski violations: Count violations and report drug-likeness assessment

Output Format

### ADMET Profile [T3]

#### Absorption
| Property | Value | Interpretation |
|----------|-------|----------------|
| BBB Penetrance | Yes | Crosses blood-brain barrier |
| Bioavailability (F20%) | 85% | Good oral absorption |

#### Distribution
| Property | Value | Interpretation |
|----------|-------|----------------|
| VDss | 1.2 L/kg | Moderate tissue distribution |
| PPB | 92% | Highly protein bound |

#### Metabolism
| CYP Enzyme | Substrate | Inhibitor |
|------------|-----------|-----------|
| CYP1A2 | No | No |
| CYP2C9 | Yes | No |
| CYP2C19 | No | No |
| CYP2D6 | No | No |
| CYP3A4 | Yes | Yes (DDI risk) |

#### Excretion
| Property | Value | Interpretation |
|----------|-------|----------------|
| Clearance | 8.5 mL/min/kg | Moderate clearance |
| Half-life | 6.2 h | Moderate half-life |

Phase 3: Toxicogenomics (CTD)

When: Compound name is resolved

Objective: Map chemical-gene-disease relationships from curated CTD data

Tools Used

Workflow

1. Call CTD_get_chemical_gene_interactions(input_terms=compound_name)
2. Call CTD_get_chemical_diseases(input_terms=compound_name)
3. Parse gene interactions: extract gene symbols, interaction types (increases/decreases expression, binding, etc.)
4. Parse disease associations: extract disease names, evidence types (marker/mechanism/therapeutic)
5. Identify most affected biological processes from gene list

Decision Logic

• Direct evidence vs inferred: CTD separates curated direct evidence from inferred associations
• Therapeutic vs toxic: Disease associations can be therapeutic (drug treats disease) or adverse (chemical causes disease)
• Gene interaction types: Distinguish between expression changes, binding, and activity modulation
• Prioritize marker/mechanism: These indicate stronger causal evidence than simple associations
• Grade curated as [T2]: Direct curated CTD evidence from literature
• Grade inferred as [T3]: Computationally inferred associations

Output Format

### Toxicogenomics (CTD) [T2/T3]

#### Chemical-Gene Interactions (Top 20)
| Gene | Interaction | Type | Evidence |
|------|------------|------|----------|
| CYP1A2 | increases expression | mRNA | [T2] curated |
| TP53 | affects activity | protein | [T2] curated |
| ...  | ... | ... | ... |

**Total interactions found**: 156
**Top affected pathways**: Xenobiotic metabolism, Apoptosis, DNA damage response

#### Chemical-Disease Associations (Top 10)
| Disease | Association Type | Evidence |
|---------|-----------------|----------|
| Liver Neoplasms | marker/mechanism | [T2] curated |
| Contact Dermatitis | therapeutic | [T2] curated |
| ... | ... | ... |

Phase 4: Regulatory Safety (FDA Labels)

When: Compound has an approved drug name

Objective: Extract regulatory safety information from FDA drug labels

Tools Used

Workflow

1. Call all 6 FDA tools in parallel (independent queries by drug name)
2. Parse and structure each response
3. Prioritize: Boxed Warnings > Contraindications > Warnings > Adverse Reactions
4. All FDA label data is [T1] evidence (regulatory finding based on human/animal data)

Decision Logic

• Boxed warning present: Flag as CRITICAL safety concern in executive summary
• No FDA data: Chemical may not be an approved drug; note "Not an FDA-approved drug" and continue with other phases
• Multiple warnings: Categorize by organ system (hepatic, cardiac, renal, CNS, etc.)
• Nonclinical toxicology: Grade as [T2] (animal data supporting human risk)

Output Format

### Regulatory Safety (FDA) [T1]

#### Boxed Warning
**PRESENT** - Hepatotoxicity risk with doses >4g/day. Liver failure reported. [T1]

#### Contraindications
- Severe hepatic impairment [T1]
- Known hypersensitivity [T1]

#### Adverse Reactions (by frequency)
| Reaction | Frequency | Severity |
|----------|-----------|----------|
| Nausea | Common (>1%) | Mild |
| Hepatotoxicity | Rare (<0.1%) | Severe |
| ... | ... | ... |

#### Nonclinical Toxicology [T2]
- **Carcinogenicity**: No carcinogenic potential in 2-year rat/mouse studies
- **Mutagenicity**: Negative in Ames assay and in vivo micronucleus test
- **Fertility**: No effects on fertility at doses up to 10x human dose

Phase 5: Drug Safety Profile (DrugBank)

When: Compound is a known drug

Objective: Retrieve curated drug safety data from DrugBank

Tools Used

Workflow

1. Call drugbank_get_safety_by_drug_name_or_drugbank_id(query=drug_name, case_sensitive=False, exact_match=False, limit=5)
2. Parse toxicity information, overdose data, contraindications
3. Cross-reference with FDA data from Phase 4

Decision Logic

• Toxicity field: Contains LD50 values, overdose symptoms, organ toxicity data
• DrugBank ID: Note if found for cross-referencing
• Conflict with FDA: If DrugBank and FDA disagree, note discrepancy and defer to FDA [T1]
• Not found: Chemical may not be in DrugBank; continue with other phases

Phase 6: Chemical-Protein Interactions (STITCH)

When: Compound can be identified by name or SMILES

Objective: Map chemical-protein interaction network for off-target assessment

Tools Used

Workflow

1. Resolve compound: STITCH_resolve_identifier(identifier=compound_name, species=9606)
2. Get interactions: STITCH_get_chemical_protein_interactions(identifiers=[stitch_id], species=9606, required_score=700)
3. Identify off-target proteins (not the intended drug target)
4. Flag safety-relevant targets: hERG (cardiac), CYP enzymes (metabolism), nuclear receptors (endocrine)

Decision Logic

• High confidence (>900): Well-established interaction [T2]
• Medium confidence (700-900): Probable interaction [T3]
• Low confidence (400-700): Possible interaction, needs validation [T4]
• Safety-relevant targets: Flag interactions with known safety targets
• No STITCH data: Chemical may be too novel; note and continue

Phase 7: Structural Alerts (ChEMBL)

When: ChEMBL molecule ID is available (from Phase 0)

Objective: Check for known toxic substructures

Tools Used

Workflow

1. If ChEMBL ID available: ChEMBL_search_compound_structural_alerts(molecule_chembl_id=chembl_id, limit=20)
2. Parse alert types: PAINS (pan-assay interference), Brenk (medicinal chemistry), Glaxo (GSK structural alerts)
3. Categorize severity: Some alerts are informational, others indicate likely toxicity

Decision Logic

• PAINS alerts: May cause false positives in screening; note for medicinal chemistry
• Brenk alerts: Known problematic substructures; flag if present
• No alerts: Good sign but not definitive proof of safety
• No ChEMBL ID: Skip this phase gracefully; note "structural alert analysis not available"

Synthesis: Integrated Risk Assessment (MANDATORY)

Always the final section. Integrates all evidence into actionable risk classification.

Risk Classification Matrix

Synthesis Template

## Integrated Risk Assessment

### Overall Risk Classification: [HIGH]

### Evidence Summary
| Dimension | Finding | Evidence Tier | Concern |
|-----------|---------|--------------|---------|
| ADMET Toxicity | DILI active, hERG active | [T3] | HIGH |
| FDA Label | Boxed warning for hepatotoxicity | [T1] | CRITICAL |
| CTD Toxicogenomics | 156 gene interactions, liver neoplasms | [T2] | HIGH |
| DrugBank | Known hepatotoxicity at high doses | [T2] | HIGH |
| STITCH | Binds CYP3A4, hERG | [T3] | MEDIUM |
| Structural Alerts | 2 Brenk alerts | [T3] | MEDIUM |

### Key Safety Concerns
1. **Hepatotoxicity** [T1]: FDA boxed warning + ADMET-AI DILI prediction + CTD liver disease associations
2. **Cardiac Risk** [T3]: ADMET-AI hERG prediction + STITCH hERG interaction
3. **Drug Interactions** [T3]: CYP3A4 substrate/inhibitor, potential DDI risk

### Data Gaps
- [ ] No in vivo genotoxicity data available
- [ ] STITCH interaction scores moderate (700-900)
- [ ] No environmental exposure data

### Recommendations
1. Avoid doses >4g/day (hepatotoxicity threshold) [T1]
2. Monitor liver function in chronic use [T1]
3. Screen for CYP3A4 interactions before co-administration [T3]
4. Consider cardiac monitoring for at-risk patients [T3]

Mandatory Completeness Checklist

Before finalizing any report, verify:

• Phase 0: Compound fully disambiguated (SMILES + CID at minimum)
• Phase 1: At least 5 toxicity endpoints reported or "prediction unavailable" noted
• Phase 2: ADMET profile with A/D/M/E sections or "not available" noted
• Phase 3: CTD queried; gene interactions and disease associations reported or "no data in CTD"
• Phase 4: FDA labels queried; results or "not an FDA-approved drug" noted
• Phase 5: DrugBank queried; results or "not found in DrugBank" noted
• Phase 6: STITCH queried; results or "no STITCH data available" noted
• Phase 7: Structural alerts checked or "ChEMBL ID not available" noted
• Synthesis: Risk classification provided with evidence summary
• Evidence Grading: All findings have [T1]-[T4] annotations
• Data Gaps: Explicitly listed in synthesis section

Tool Parameter Reference

Critical Parameter Notes (verified from source code):

Response Format Notes

• ADMET-AI: Returns {status: "success", data: {...}} with prediction values
• CTD: Returns list of interaction/association objects
• FDA: Returns {status, data} with label text
• DrugBank: Returns {data: [...]} with drug records
• STITCH: Returns list of interaction objects with scores
• PubChem CID lookup: Returns {IdentifierList: {CID: [...]}} (may or may not have data wrapper)
• PubChem properties: Returns dict with CID, MolecularWeight, ConnectivitySMILES, IUPACName

Fallback Strategies

Compound Resolution

• Primary: PubChem by name -> CID -> properties -> SMILES
• Fallback 1: ChEMBL search by name -> molecule -> SMILES
• Fallback 2: If SMILES provided directly, skip name resolution

Toxicity Prediction

• Primary: All 9 ADMET-AI endpoints
• Fallback: If ADMET-AI fails for a compound, note "prediction failed" and continue with database evidence
• Note: ADMET-AI may fail for very large or unusual SMILES

Regulatory Data

• Primary: FDA labels by drug name
• Fallback: If FDA returns no data, try alternative drug names (brand vs generic)
• Note: Non-drug chemicals (pesticides, industrial) will not have FDA labels

CTD Data

• Primary: Search by common chemical name
• Fallback: Try MeSH name if common name fails
• Note: Novel compounds may not be in CTD

Common Use Patterns

Pattern 1: Novel Compound Assessment

Input: SMILES string for new molecule
Workflow: Phase 0 (SMILES->CID) -> Phase 1 (toxicity) -> Phase 2 (ADMET) -> Phase 7 (structural alerts) -> Synthesis
Output: Predictive safety profile for novel compound

Pattern 2: Approved Drug Safety Review

Input: Drug name (e.g., "Acetaminophen")
Workflow: All phases (0-7 + Synthesis)
Output: Complete safety dossier with regulatory + predictive + database evidence

Pattern 3: Environmental Chemical Risk

Input: Chemical name (e.g., "Bisphenol A")
Workflow: Phase 0 -> Phase 1 -> Phase 2 -> Phase 3 (CTD, key for env chemicals) -> Phase 6 -> Synthesis
Output: Environmental health risk assessment focused on gene-disease associations

Pattern 4: Batch Toxicity Screening

Input: Multiple SMILES strings
Workflow: Phase 0 -> Phase 1 (batch) -> Phase 2 (batch) -> Comparative table -> Synthesis
Output: Comparative toxicity table ranking compounds by safety

Pattern 5: Toxicogenomic Deep-Dive

Input: Chemical name + specific gene or disease interest
Workflow: Phase 0 -> Phase 3 (CTD expanded) -> Literature search -> Synthesis
Output: Detailed chemical-gene-disease mechanistic analysis

Output Report Structure

All analyses generate a structured markdown report with progressive sections:

# Chemical Safety & Toxicology Report: [Compound Name]

**Generated**: YYYY-MM-DD HH:MM
**Compound**: [Name] | SMILES: [SMILES] | CID: [CID]

## Executive Summary
[2-3 sentence overview with risk classification and key findings, all graded]

## 1. Compound Identity
[Phase 0 results - disambiguation table]

## 2. Predictive Toxicology
[Phase 1 results - ADMET-AI toxicity endpoints]

## 3. ADMET Profile
[Phase 2 results - absorption, distribution, metabolism, excretion]

## 4. Toxicogenomics
[Phase 3 results - CTD chemical-gene-disease relationships]

## 5. Regulatory Safety
[Phase 4 results - FDA label information]

## 6. Drug Safety Profile
[Phase 5 results - DrugBank data]

## 7. Chemical-Protein Interactions
[Phase 6 results - STITCH network]

## 8. Structural Alerts
[Phase 7 results - ChEMBL alerts]

## 9. Integrated Risk Assessment
[Synthesis - risk classification, evidence summary, data gaps, recommendations]

## Appendix: Methods and Data Sources
[Tool versions, databases queried, date of access]

Limitations & Known Issues

Tool-Specific

• ADMET-AI: Predictions are computational [T3]; should not replace experimental testing
• CTD: Curated but may lag behind latest literature by 6-12 months
• FDA: Only covers FDA-approved drugs; not applicable to environmental chemicals or supplements
• DrugBank: Primarily drugs; limited coverage of industrial chemicals
• STITCH: Score thresholds affect sensitivity; lower scores increase false positives
• ChEMBL: Structural alerts require ChEMBL ID; not all compounds have one

Analysis

• Novel compounds: May only have ADMET-AI predictions (no database evidence)
• Environmental chemicals: FDA/DrugBank phases will be empty; rely on CTD and ADMET-AI
• Batch mode: ADMET-AI can handle batches; other tools require individual queries
• Species specificity: Most data is human-centric; animal data noted where applicable

Technical

• SMILES validity: Invalid SMILES will cause ADMET-AI failures
• Name ambiguity: Chemical names can be ambiguous; always verify with CID
• Rate limits: Some FDA endpoints may rate-limit for rapid queries

Summary

Chemical Safety & Toxicology Assessment Skill provides comprehensive safety evaluation by integrating:

1. Predictive toxicology (ADMET-AI) - 9 tools covering toxicity, ADMET, physicochemical properties
2. Toxicogenomics (CTD) - Chemical-gene-disease relationship mapping
3. Regulatory safety (FDA) - 6 tools for label-based safety extraction
4. Drug safety (DrugBank) - Curated toxicity and contraindication data
5. Chemical interactions (STITCH) - Chemical-protein interaction networks
6. Structural alerts (ChEMBL) - Known toxic substructure detection

Outputs: Structured markdown report with risk classification, evidence grading, and actionable recommendations

Best for: Drug safety assessment, chemical hazard profiling, environmental toxicology, ADMET characterization, toxicogenomic analysis

Total tools integrated: 25+ tools across 6 databases