ADMET Prediction (ToolUniverse Claude Skill)

A ToolUniverse agent skill that turns a drug name or SMILES into a structured ADMET scorecard, combining machine-learned predictions with experimental toxicity records.

   
Type Claude Skill (one of ToolUniverse’s pre-built agent skills)
Supplier Zitnik Lab, Harvard Medical School
Availability GA — part of the ToolUniverse skills collection (skills/tooluniverse-admet-prediction/)
Pricing Free / OSS (Apache-2.0); wraps public services (PubChem, PubChemTox, ChEMBL, ADMET-AI, SwissADME)
Capabilities Read-only — drives ToolUniverse tool calls; no data writes
Verified works · 2026-08-06
Security cleared · 2026-08-06 — ToolUniverse Apache-2.0, public read-only APIs, no credentials

How to install

This skill calls ToolUniverse tools, so the ToolUniverse MCP server must be installed first (see the ToolUniverse page). Simplest registration:

claude mcp add --transport stdio tooluniverse -- uvx tooluniverse

Then add the skills:

  • Claude Code — install the whole skill collection (the skill resolves as tooluniverse-admet-prediction): 
    npx skills add mims-harvard/ToolUniverse
  • Manual / other agents — copy just this skill directory into your skills folder: 
    git clone https://github.com/mims-harvard/ToolUniverse
cp -r ToolUniverse/skills/tooluniverse-admet-prediction ~/.claude/skills/

    (replace ~/.claude/skills/ with your agent’s skills directory if you are not using Claude Code/Desktop.)

The skill sets disable-model-invocation: true upstream, so invoke it explicitly (e.g. ask Claude to “use the admet-prediction skill”) rather than relying on automatic dispatch.

What it does

Runs a five-phase profiling workflow:

  1. Identity resolution — resolve a drug name to a PubChem CID and canonical SMILES before any prediction runs.
  2. Physicochemical and drug-likeness — Lipinski rule-of-five compliance, TPSA, logP, solubility, and PAINS/Brenk structural alerts via ADMET-AI and SwissADME.
  3. ADME prediction — blood-brain-barrier penetrance, oral bioavailability, CYP450 interactions, clearance, and distribution.
  4. Toxicity assessment — predicted AMES mutagenicity, drug-induced liver injury (DILI), hERG cardiotoxicity, and carcinogenicity, cross-checked against experimental PubChemTox records (LD50/LC50, GHS classification, target organs) plus nuclear-receptor and stress-response assay panels.
  5. Scorecard — aggregates into a 13-category scorecard with pass/warn/fail verdicts, evidence tiers (T1–T3), and recommended next steps; ChEMBL supplies clinical phase and regulatory context.

Operating rules the skill enforces: identity is resolved before analysis, experimental data outranks prediction, and a toxicity FAIL on hERG, AMES, or DILI is treated as program-limiting until wet-lab evidence refutes it.

Primary use cases: pre-screening a compound library before assay spend, drug-likeness triage, CNS-penetrance and hepatotoxicity risk assessment.

Notes

It is a reasoning layer over ToolUniverse; without the MCP server registered, the tool calls fail. Every value in the scorecard other than the PubChemTox and ChEMBL rows is a model prediction, not a measurement — the evidence tiers exist so a reader can tell which is which.

Overlaps with the ADMET phase of the Small Molecule Discovery skill but goes considerably deeper on toxicity and evidence grading; use that skill for end-to-end compound characterization and this one when ADMET is the question. For a self-hostable MCP alternative built on ADMETlab 3.0 rather than ADMET-AI, see ADMETlab MCP Server; Inductive Bio is the enterprise-gated connector in the same space. ToolUniverse ships ~68 such skills; other workflows are catalogued separately.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.