Skip to main content
skill CX Codex

tooluniverse-admet-prediction — Claude Code skill for ADMET profiling

The tooluniverse-admet-prediction Claude Code skill profiles drug candidates by ADMET properties. It helps discard compounds likely to fail before wet-lab work.

Apache-2.0 245 installs ★ 1,798 updated 2026-10-24

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-admet-prediction -a claude-code
Open on GitHub

What tooluniverse-admet-prediction does

The skill pulls together ADMET-AI predictions, SwissADME drug-likeness checks, PubChemTox experimental toxicity, ChEMBL clinical data, Lipinski's rule of five, and CYP interaction records. Starting from a drug name or a SMILES string, it runs five phases: identity resolution through PubChem, physicochemical and drug-likeness checks, ADME predictions, toxicity assessment, and a final pass/warn/fail scorecard with clinical context.

Its core rule is compute, don't describe. When an analysis needs statistics or scoring, the agent writes Python and runs it through Bash instead of speculating. Each property is judged on its own: strong absorption does not offset liver toxicity. Experimental data (T2) outranks computational predictions (T3), and any toxicity FAIL (hERG, AMES, DILI) is treated as program-limiting until wet-lab data says otherwise.

The ADMETAI tools need the tooluniverse[ml] extra: uv pip install 'tooluniverse[ml]'. Without it, tools still show up in tu list but fail at call time with ADMETModel requires 'admet-ai' package. The first ADMETAI call loads PyTorch and prints GPU warnings and TypedStorage deprecation notices — that is library noise during CPU inference, not a failure. Only an error field or missing predictions count as a real problem.

What it helps with

  • check whether a compound violates Lipinski's rule of five
  • assess blood-brain barrier penetration from a SMILES string
  • flag hepatotoxicity and hERG risk before lab testing
  • build an ADME/PK profile from a drug name or SMILES
  • screen compound libraries and drop molecules with toxicity FAILs

Example prompts for the agent

  • What are the ADMET properties of ibuprofen?
  • Will this compound cross the blood-brain barrier: CC(C)Cc1ccc(cc1)C(C)C(=O)O
  • Check this molecule against Lipinski's rule of five

How to install tooluniverse-admet-prediction

1Skills CLI (Vercel)

npx skills add mims-harvard/ToolUniverse --skill tooluniverse-admet-prediction -a claude-code

2Manually: put the files in place

~/.claude/skills/tooluniverse-admet-prediction/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/tooluniverse-admet-prediction/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about tooluniverse-admet-prediction

How do I install tooluniverse-admet-prediction in Claude Code?

Run npx skills add mims-harvard/ToolUniverse --skill tooluniverse-admet-prediction -a claude-code. The skill lands in ~/.claude/skills/tooluniverse-admet-prediction/SKILL.md.

What does the skill require to run?

ADMETAI tools run a local model and need tooluniverse[ml]: uv pip install 'tooluniverse[ml]'. Calls fail with ADMETModel requires 'admet-ai' package otherwise. Run tooluniverse-doctor to see which optional groups are installed.

How does it differ from a plain agent prompt?

It forces the agent to resolve compound identity through PubChem first instead of guessing SMILES or CID values. ADMETAI tools expect smiles: list[str], SwissADME takes a single string, PubChemTox takes a CID or name. Every analysis ends with a graded pass/warn/fail scorecard.

Source text

SKILL.md · Apache-2.0
---
name: tooluniverse-admet-prediction
description: Comprehensive ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) profiling for drug candidates. Integrates ADMET-AI predictions, SwissADME drug-likeness, PubChemTox experimental toxicity, ChEMBL clinical data, Lipinski rule-of-five, and CYP interaction data. Use for drug-likeness assessment, BBB penetration, bioavailability, hepatotoxicity prediction, ADME/PK profiling, or screening compound libraries before lab testing.
disable-model-invocation: true
---

# ADMET Prediction & Drug Candidate Profiling

**ADMET reasoning**: a drug fails if it can't be absorbed, distributes to wrong tissues, isn't metabolized safely, or isn't excreted. Evaluate each property independently — good absorption doesn't compensate for liver toxicity. The ADME properties determine whether a compound reaches its target at therapeutic concentrations; toxicity determines whether it's safe to do so. Prioritize experimental data (T2) over computational predictions (T3) — ADMETAI predictions are screening tools, not definitive verdicts. When a FAIL is flagged in any toxicity category (hERG, AMES, DILI), treat it as program-limiting until wet-lab data refutes it.

**LOOK UP DON'T GUESS**: never assume SMILES, CID, or experimental LD50 values — always call PubChem to resolve compound identity before any ADMETAI or PubChemTox call.

Comprehensive pharmacokinetic and toxicity profiling integrating AI-based ADMET predictions, rule-based drug-likeness filters, and experimental benchmarks from curated databases.

## When to Use This Skill

**Triggers**:
- "What are the ADMET properties of [compound]?"
- "Is [drug] likely to cross the blood-brain barrier?"
- "Predict the toxicity of this SMILES: ..."
- "Does [compound] violate Lipinski's rule of five?"
- "Assess the drug-likeness of [molecule]"
- "What are the CYP interactions for [drug]?"
- "Pharmacokinetic profile of [compound]"
- "Is [compound] orally bioavailable?"
- "What is the LD50 / hERG liability of [molecule]?"

**Input**: Drug name (e.g., "ibuprofen") OR SMILES string (e.g., "CC(C)Cc1ccc(cc1)C(C)C(=O)O")

## Before You Run

ADMETAI tools run a local model, so they need the `ml` extra:

```bash
uv pip install 'tooluniverse[ml]'
```

Without it the tools still appear in `tu list` (the config loads) but fail at
call time with `ADMETModel requires 'admet-ai' package`. Run
`tooluniverse-doctor` to confirm which optional groups are installed.

**Expected console noise — not errors.** The first ADMETAI call loads PyTorch
and prints warnings such as missing-GPU / `Trainer` messages from
PyTorch Lightning, and `TypedStorage is deprecated` from PyTorch. These are
emitted by the underlying libraries during normal CPU inference. Predictions
are unaffected — do not report them to the user as failures and do not retry
the call because of them. Only treat output as a failure if the tool returns an
`error` field or no predictions.

---

## 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. **Resolve identity first** - Always convert drug name to SMILES before calling ADMETAI tools
2. **ADMETAI tools require `tooluniverse[ml]`** - If import fails, skip to SwissADME/PubChemTox fallbacks
3. **All ADMETAI tools take `smiles: list[str]`** - Always wrap in a list, even for one compound
4. **SwissADME takes `smiles: str`** - Single string, NOT a list (SOAP-style with `operation` param)
5. **PubChemTox tools accept `cid` or `compound_name`** - Use CID when available for reliability
6. **Evidence grading mandatory** - Predictions (T3), experimental data (T2), regulatory (T1)
7. **Scorecard output** - Every analysis must end with a pass/warn/fail scorecard
8. **Explain significance** - State WHY each property matters for drug development

---

## Evidence Grading

| Tier | Label | Source |
|------|-------|--------|
| **T1** | Regulatory/Clinical | FDA labels, ChEMBL max clinical phase |
| **T2** | Experimental | PubChemTox LD50/LC50, in vitro AMES, animal studies |
| **T3** | Computational | ADMETAI predictions, SwissADME calculations |
| **T4** | Annotation | Database cross-references, text-mined |

## Workflow: 5-Phase ADMET Profiling

```
User Query (drug name or SMILES)
|
+-- PHASE 1: Compound Identity Resolution
|   PubChem name->CID->SMILES, or validate input SMILES
|
+-- PHASE 2: Physicochemical & Drug-Likeness
|   ADMETAI physicochemical + SwissADME druglikeness -> Lipinski/Veber
|
+-- PHASE 3: ADME Predictions
|   BBB, bioavailability, CYP interactions, clearance, solubility
|
+-- PHASE 4: Toxicity Assessment
|   ADMETAI tox + PubChemTox experimental + nuclear receptor + stress
|
+-- PHASE 5: Scorecard & Clinical Context
|   ChEMBL max phase, aggregate pass/warn/fail, final recommendation
```

---

### PHASE 1: Compound Identity Resolution

**Goal**: Obtain SMILES, PubChem CID, and basic identifiers for the query compound.

**Steps**:

1. **If input is a drug name**:
   - Call `PubChem_get_CID_by_compound_name(name=<drug_name>)` to get CID
   - Call `PubChem_get_compound_properties_by_CID(cid=<CID>)` to get SMILES and MW
   - Extract `ConnectivitySMILES` from the response (NOT `CanonicalSMILES`)

2. **If input is a SMILES string**:
   - Call `PubChem_get_CID_by_SMILES(smiles=<SMILES>)` to get CID
   - Call `PubChem_get_compound_properties_by_CID(cid=<CID>)` for compound name and MW
   - Use the input SMILES for all subsequent ADMETAI calls

3. **Record**:
   - Compound name, CID, SMILES, molecular formula, molecular weight, IUPAC name
   - If CID lookup fails, proceed with SMILES only (ADMETAI does not need CID)

**Why this matters**: ADMETAI tools require SMILES input. PubChemTox tools work best with CID. Resolving both ensures all downstream tools can be called. PubChem is the authoritative source for SMILES canonicalization.

**Fallback**: If PubChem has no entry, the user must provide SMILES directly. Cannot proceed without SMILES.

---

### PHASE 2: Physicochemical Properties & Drug-Likeness

**Goal**: Evaluate whether the compound has drug-like physicochemical properties.

**Steps**:

1. **ADMETAI physicochemical** (primary):
   ```
   ADMETAI_predict_physicochemical_properties(smiles=["<SMILES>"])
   ```
   Returns: MW, logP, TPSA, HBD, HBA, rotatable bonds

2. **SwissADME drug-likeness** (complementary):
   ```
   SwissADME_check_druglikeness(operation="check_druglikeness", smiles="<SMILES>")
   SwissADME_calculate_adme(operation="calculate_adme", smiles="<SMILES>")
   ```
   Returns: Lipinski, Veber, Ghose, Egan, Muegge rule compliance; PAINS alerts; Brenk alerts

3. **ADMETAI solubility**:
   ```
   ADMETAI_predict_solubility_lipophilicity_hydration(smiles=["<SMILES>"])
   ```
   Returns: Aqueous solubility (LogS), lipophilicity, hydration free energy

**Interpret & Score**:

| Property | Ideal Range | Why It Matters |
|----------|-------------|----------------|
| MW | < 500 Da | Larger molecules have poor membrane permeability (Lipinski) |
| LogP | -0.4 to 5.6 | Too hydrophobic = poor solubility; too hydrophilic = poor permeability |
| HBD | <= 5 | Excess donors reduce membrane crossing (Lipinski) |
| HBA | <= 10 | Excess acceptors reduce membrane crossing (Lipinski) |
| TPSA | < 140 A^2 | High PSA correlates with poor oral absorption |
| Rotatable bonds | <= 10 | Molecular flexibility affects bioavailability (Veber) |
| LogS | > -6 | Below -6 = practically insoluble, formulation challenge |
| PAINS alerts | 0 | Pan-assay interference compounds give false positives in screens |

**Verdict**: PASS if Lipinski <= 1 violation and no PAINS alerts; WARN if 2 violations; FAIL if 3+ violations or PAINS+.

**Fallback**: If ADMETAI import fails (missing `tooluniverse[ml]`), rely on SwissADME alone. SwissADME provides all Lipinski descriptors independently.

**Local TDC oracle scoring (no ML extra, no PubChem round-trip needed)**: `TDC_predict_oracle_score(smiles=..., oracle=...)` computes drug-likeness/property scores locally via the PyTDC package straight from a SMILES string — useful when `tooluniverse[ml]` isn't installed or as a fast independent cross-check. Six oracles: `QED` (drug-likeness 0-1), `SA` (synthetic accessibility 1=easy to 10=hard), `LogP`, and three target-binding-likelihood oracles (`GSK3B`, `JNK3`, `DRD2`) useful for early candidate triage against those specific targets. **Verified live and cross-checked**: `TDC_predict_oracle_score(smiles="CC(=O)Oc1ccccc1C(=O)O", oracle="SA")` and the independent `Chem_sa_score(smiles=...)` tool both returned **1.580** for aspirin — same underlying RDKit SA_Score algorithm, two separate wrapper tools; either is fine to cite, don't report both as independent corroborating evidence since they're the same computation.

`TDC_load_dataset(problem=..., name=...)` / `TDC_list_datasets(problem=...)` give access to TDC's ADME/Tox/HTS benchmark datasets — relevant if validating a custom prediction model against a public benchmark, not part of a single-compound ADMET profile. **`TDC_list_datasets` works** (verified live: `problem="ADME"` → 27 real dataset names). **`TDC_load_dataset` is currently broken** — verified live and reproduced on two different datasets (`Caco2_Wang`, `hERG`): it downloads successfully but the loader then receives an HTML error page instead of data and raises `Please report this error to [email protected]`. Don't build a workflow around it; use `TDC_list_datasets` for discovery and note the load failure if a user specifically needs a dataset pulled.

---

### PHASE 3: ADME Predictions

**Goal**: Predict absorption, distribution, metabolism, and excretion behavior.

**Steps**:

1. **Blood-brain barrier penetration**:
   ```
   ADMETAI_predict_BBB_penetrance(smiles=["<SMILES>"])
   ```
   - BBB+ = compound can cross; BBB- = cannot
   - Critical for CNS drugs (must cross) and peripherally-acting drugs (should NOT cross to avoid CNS side effects)

2. **Oral bioavailability**:
   ```
   ADMETAI_predict_bioavailability(smiles=["<SMILES>"])
   ```
   - F20% = at least 20% oral bioavailability; F30% = at least 30%
   - Low bioavailability means the drug is extensively metabolized or poorly absorbed
   - F < 20% generally requires non-oral routes (IV, inhaled, topical)

3. **CYP450 interactions**:
   ```
   ADMETAI_predict_CYP_interactions(smiles=["<SMILES>"])
   ```
   - Reports substrate/inhibitor status for CYP1A2, 2C9, 2C19, 2D6, 3A4
   - **Why CYP matters**: ~75% of drugs are metabolized by CYP enzymes. Inhibiting CYP3A4 (which metabolizes ~50% of drugs) causes dangerous drug-drug interactions (DDIs). CYP2D6 polymorphisms affect ~25% of drugs -- poor metabolizers accumulate toxic levels
   - Substrate of CYP2D6 = pharmacogenomic risk (poor/ultra-rapid metabolizers)
   - Inhibitor of CYP3A4 = high DDI risk (co-administered drugs accumulate)

4. **Clearance and distribution**:
   ```
   ADMETAI_predict_clearance_distribution(smiles=["<SMILES>"])
   ```
   - VDss (volume of distribution): low (<0.7 L/kg) = confined to plasma; high (>1 L/kg) = distributed to tissues
   - Clearance: high clearance = short half-life, frequent dosing needed
   - Plasma protein binding (PPB): >95% bound = narrow therapeutic window, DDI risk from displacement

5. **SwissADME pharmacokinetics** (cross-validation):
   - GI absorption (high/low), P-gp substrate status, skin permeation (logKp)

**Key flags**: BBB+ for non-CNS drug (WARN: CNS side effects); BBB- for CNS drug (FAIL: won't reach target); F < 20% (WARN: poor oral bioavailability); CYP3A4 inhibitor (WARN: high DDI); CYP2D6 substrate (WARN: pharmacogenomic variability); PPB > 99% (WARN: narrow window); high clearance + low bioavailability (FAIL).

**Fallback**: If ADMETAI unavailable, SwissADME provides GI absorption, BBB permeation (yes/no), P-gp substrate, and CYP inhibition predictions.

---

### PHASE 4: Toxicity Assessment

**Goal**: Evaluate safety liabilities from both predicted and experimental sources.

**Steps**:

1. **ADMETAI toxicity predictions** [T3]:
   ```
   ADMETAI_predict_toxicity(smiles=["<SMILES>"])
   ```
   Key endpoints:
   - **AMES**: Mutagenicity (bacterial reverse mutation test). Positive = potential carcinogen; regulatory agencies require AMES testing for all new drugs
   - **DILI**: Drug-induced liver injury risk. Leading cause of drug withdrawal (e.g., troglitazone). Positive = hepatotoxicity concern requiring liver function monitoring
   - **hERG**: hERG potassium channel inhibition. Causes QT prolongation and fatal cardiac arrhythmia. hERG+ = cardiotoxicity liability; multiple drugs withdrawn for this (e.g., terfenadine, cisapride)
   - **ClinTox**: Clinical trial toxicity / FDA withdrawal risk. Trained on drugs that failed trials or were withdrawn for toxicity
   - **LD50_Zhu**: Predicted lethal dose (mg/kg, rat oral). Lower = more acutely toxic
   - **Skin_Reaction**: Dermal sensitization potential. Important for topical drugs
   - **Carcinogens_Lagunin**: Carcinogenicity prediction

2. **Nuclear receptor activity** [T3]:
   ```
   ADMETAI_predict_nuclear_receptor_activity(smiles=["<SMILES>"])
   ```
   - AR (androgen receptor), ER (estrogen receptor), AhR, PPAR-gamma activity
   - Positive = potential endocrine disruption; critical for chronic-use drugs and environmental chemicals

3. **Stress response pathways** [T3]:
   ```
   ADMETAI_predict_stress_response(smiles=["<SMILES>"])
   ```
   - p53 activation = DNA damage response (genotoxicity signal)
   - MMP disruption = mitochondrial toxicity
   - ATAD5 = DNA repair stress
   - HSE = heat shock / protein misfolding stress

4. **PubChemTox experimental data** [T2] (call all in parallel):
   ```
   PubChemTox_get_toxicity_values(cid=<CID>)
   PubChemTox_get_ghs_classification(cid=<CID>)
   PubChemTox_get_acute_effects(cid=<CID>)
   PubChemTox_get_carcinogen_classification(cid=<CID>)
   PubChemTox_get_target_organs(cid=<CID>)
   PubChemTox_get_toxicity_summary(cid=<CID>)
   ```
   - Real animal study data (LD50, LC50, NOAEL) anchors computational predictions
   - GHS classification provides internationally harmonized hazard categories
   - Carcinogen classification from IARC (Group 1/2A/2B), NTP, EPA

**Key flags**: AMES positive (FAIL: mutagenic); DILI positive (WARN: hepatotox); hERG positive (FAIL: cardiac, often program-killing); ClinTox positive (WARN); LD50 < 50 mg/kg (FAIL: GHS 1-2); LD50 50-300 mg/kg (WARN: GHS 3); NR-ER/AR active (WARN: endocrine disruption); p53 active (WARN: genotoxicity); IARC Group 1/2A (FAIL: known/probable carcinogen).

**Fallback**: If ADMETAI unavailable, PubChemTox provides experimental toxicity data for known compounds. For novel compounds without PubChem entries, flag as "no experimental toxicity data available -- computational predictions only."

---

### PHASE 5: Scorecard Assembly & Clinical Context

**Goal**: Aggregate all findings into a structured ADMET scorecard with pass/warn/fail verdicts.

**Steps**:

1. **ChEMBL clinical status** [T1] (if drug has ChEMBL ID):
   ```
   ChEMBL_get_molecule(chembl_id="<CHEMBL_ID>")
   ```
   - Max phase: 4 = approved, 3 = Phase III, 2 = Phase II, 1 = Phase I, 0 = preclinical
   - Ro5 violations from ChEMBL (independent validation of Lipinski)
   - First approval year, indication class, black box warning flag

2. **Build the ADMET Scorecard**: produce a table with 13 categories (Physicochemical, Solubility, Absorption, Distribution, Metabolism, Excretion, Tox: Mutagenicity/Hepatotoxicity/Cardiotoxicity/Carcinogenicity/Acute, Endocrine, Clinical Tox), each with PASS/WARN/FAIL verdict and key finding. Include compound identity header and overall verdict. Tag each finding with evidence tier [T1-T3].

3. **Interpretation narrative**: After the scorecard, provide a 3-5 sentence summary:
   - Highlight the most critical findings (any FAILs or WARNs)
   - State whether the compound is suitable for oral administration
   - Note any DDI risks from CYP interactions
   - Flag pharmacogenomic concerns (CYP2D6 substrate)
   - Recommend next steps (e.g., "hERG patch clamp assay recommended to confirm computational prediction")

---

## Completeness Checklist (MANDATORY before reporting)

Before delivering the final scorecard, verify:

- [ ] Compound identity resolved (name, CID, SMILES all present or explicitly noted as unavailable)
- [ ] Physicochemical properties reported with Lipinski verdict
- [ ] At least one source for each ADME property (ADMETAI or SwissADME)
- [ ] All 7 ADMETAI toxicity endpoints reported (or marked N/A with reason)
- [ ] PubChemTox experimental data checked (even if "no data found")
- [ ] Nuclear receptor and stress response checked (or marked N/A)
- [ ] Evidence tier tagged for every finding
- [ ] Scorecard table complete with verdicts for all 13 categories
- [ ] Overall verdict stated
- [ ] Interpretation narrative provided with actionable next steps

© mims-harvard · Apache-2.0 · mims-harvard/ToolUniverse

Similar skills

All skills for Science and research →

From the same set136

The whole set →
tooluniverse-toxicology tooluniverse-toxicology builds a toxicity picture for a drug or chemical: AOP mechanism, FAERS signals, FDA label warnings, CTD associations, EPA CompTox hazard data. CX Codex +1 250installs tooluniverse-drug-mechanism-research The tooluniverse-drug-mechanism-research skill follows a drug from its primary target through pathway changes to clinical outcomes. Use it for pharmacology research, target comparisons, and mechanism reports. CX Codex +1 250installs tooluniverse-regulatory-genomics tooluniverse-regulatory-genomics queries regulatory elements, transcription factor binding motifs, and predicts non-coding variant effects using genomic databases and deep learning models. CX Codex +1 247installs tooluniverse-data-integration-analysis tooluniverse-data-integration-analysis connects statistical results like DEGs and GWAS hits with biological databases. Explains mechanisms behind p-values. CX Codex +1 246installs tooluniverse-data-wrangling tooluniverse-data-wrangling guides the agent to download and parse scientific datasets via Python in Bash when ToolUniverse tools return only metadata or truncate results. CX Codex +1 245installs tooluniverse-protein-structure-prediction tooluniverse-protein-structure-prediction predicts protein structures from amino acid sequences and gathers evidence to assess the models. Use it to examine a new protein or explore how a variant may affect its structure. CX Codex +1 245installs tooluniverse-chemical-sourcing tooluniverse-chemical-sourcing finds commercial chemical suppliers across ZINC, Enamine, eMolecules, and Mcule after validating structures via PubChem or ChEMBL. CX Codex +1 243installs tooluniverse-clinical-data-integration tooluniverse-clinical-data-integration builds a drug-safety review from FDA labels, FAERS, pharmacogenomics, trials, and literature. Use it for pharmacovigilance and regulatory assessment. CX Codex +1 242installs

Project sponsors

Companies that keep this analytics open