tooluniverse-chemical-safety

What the source does

• FilesystemNone observedNo reads or writes in the source
• NetworkNone observedNo outbound calls in the source
• ShellNone observedNo commands run in the source
• DependenciesPinnedNothing floats
• SecretsNone foundNo keys in source

Five audit layers

Gates applied: no behavioural pass.

Audit 0.4.1 · source sha 43be0fac747b · 31 Oct 2026Report an issue or request a re-scan

• Publishermims-harvard GitHub organizationOwner of the repository this skill is read from
• Source foldermims-harvard/tooluniverse/skills/tooluniverse-chemical-safetyThe exact folder we audited
• LicenceApache-2.0As the repository states it
• Commit pinned for this verdict43be0fac747bA new audit runs when the source changes
• Repository statusNot archivedPushed today
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Chemical safety and toxicology assessment integrating ADMET-AI predictions, CTD toxicogenomics, PubChemTox experimental data, GHS/IARC hazard classification, and exposure-context analysis. Use for chemical hazard…

---
name: tooluniverse-chemical-safety
description: Chemical safety and toxicology assessment integrating ADMET-AI predictions, CTD toxicogenomics, PubChemTox experimental data, GHS/IARC hazard classification, and exposure-context analysis. Use for chemical hazard identification, occupational/consumer-product toxicity, dose-response evaluation, and acute (LD50) vs chronic toxicity assessment. Distinguishes drug toxicity from environmental chemical toxicity.
disable-model-invocation: true
---

# Chemical Safety & Toxicology Assessment

**Toxicity assessment**: identify the chemical, check known hazards (GHS, IARC), then look for ADMET predictions. Dose makes the poison — always consider exposure level, as a compound that is toxic at high doses may be safe at relevant exposures. Distinguish between acute toxicity (LD50, GHS category) and chronic hazards (carcinogenicity, endocrine disruption) — they require different risk management approaches. Computational predictions (ADMETAI) are T3 evidence and must be anchored by experimental data from PubChemTox or FDA labels wherever available. When evidence conflicts between prediction and experiment, always defer to the experimental finding.

**LOOK UP DON'T GUESS**: never assume GHS categories, IARC classification, or CTD disease links — always call PubChemTox and CTD tools to retrieve current classifications before reporting.

Comprehensive chemical safety analysis integrating predictive AI models, curated toxicogenomics databases, regulatory safety data, and chemical-biological interaction networks.

## When to Use This Skill

**Triggers**:
- "Is this chemical toxic?" / "Assess the safety profile of [drug/chemical]"
- "What are the ADMET properties of [SMILES]?"
- "What genes does [chemical] interact with?" / "What diseases are linked to [chemical] exposure?"
- "Drug safety assessment" / "Environmental health risk" / "Chemical hazard profiling"

**Use Cases**:
1. **Predictive Toxicology**: AI-predicted endpoints (AMES, DILI, LD50, carcinogenicity, hERG) via SMILES
2. **ADMET Profiling**: Absorption, distribution, metabolism, excretion, toxicity
3. **Toxicogenomics**: Chemical-gene-disease mapping from CTD
4. **Regulatory Safety**: FDA label warnings, contraindications, adverse reactions
5. **Drug Safety**: DrugBank safety + FDA labels combined
6. **Chemical-Protein Interactions**: STITCH-based interaction networks
7. **Environmental Toxicology**: Chemical-disease associations for contaminants

---

## 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.

## 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 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)

| Tier | Symbol | Criteria | Examples |
|------|--------|----------|----------|
| **T1** | [T1] | Direct human evidence, regulatory finding | FDA boxed warning, clinical trial toxicity |
| **T2** | [T2] | Animal studies, validated in vitro | Nonclinical toxicology, AMES positive, animal LD50 |
| **T3** | [T3] | Computational prediction, association data | ADMET-AI prediction, CTD association |
| **T4** | [T4] | Database annotation, text-mined | Literature mention, unvalidated database entry |

Evidence grades MUST appear in: Executive Summary, Toxicity Predictions, Regulatory Safety, Chemical-Gene Interactions, Risk Assessment.

---

## Core Strategy: 8 Research Phases

```
Chemical/Drug Query
|
+-- PHASE 0: Compound Disambiguation (ALWAYS FIRST)
|   Resolve name -> SMILES, PubChem CID, ChEMBL ID, formula, weight
|
+-- PHASE 1: Predictive Toxicology (ADMET-AI)
|   AMES, DILI, ClinTox, carcinogenicity, LD50, hERG, skin reaction
|   Stress response pathways, nuclear receptor activity
|
+-- PHASE 2: ADMET Properties
|   BBB penetrance, bioavailability, clearance, CYP interactions, physicochemical
|
+-- PHASE 3: Toxicogenomics (CTD)
|   Chemical-gene interactions, chemical-disease associations
|
+-- PHASE 4: Regulatory Safety (FDA Labels)
|   Boxed warnings, contraindications, adverse reactions, nonclinical tox
|
+-- PHASE 5: Drug Safety Profile (DrugBank)
|   Toxicity data, contraindications, drug interactions
|
+-- PHASE 6: Chemical-Protein Interactions (STITCH)
|   Direct binding, off-target effects, interaction confidence
|
+-- PHASE 7: Structural Alerts (ChEMBL)
|   PAINS, Brenk, Glaxo structural alerts
|
+-- SYNTHESIS: Integrated Risk Assessment
    Risk classification, evidence summary, data gaps, recommendations
```

See **phase-procedures-detailed.md** for complete tool parameters, decision logic, output templates, and fallback strategies for each phase.

---

## Tool Summary by Phase

### Phase 0: Compound Disambiguation
- `PubChem_get_CID_by_compound_name` (`name`: str)
- `PubChem_get_compound_properties_by_CID` (`cid`: int)
- `ChEMBL_get_molecule` (if ChEMBL ID available)

### Phase 1: Predictive Toxicology
> **Dependency**: ADMET-AI tools require `pip install tooluniverse[ml]`. If unavailable, skip to Phase 3 and use 

Excerpt, first 6,000 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

git clone https://github.com/mims-harvard/tooluniverse
git -C tooluniverse checkout 43be0fac747b

Works in

1. 2026-10-31
   ✓SAFE B 89 / 100

   first audit · source 43be0fac747b

Questions about tooluniverse-chemical-safety

• github.com/mims-harvard/tooluniverse/tree/main/skills/tooluniver — the audited source at commit 43be0fac747b and every finding aboveObserved · 31 Oct 2026

Spotted something wrong? Report a correction · Methodology