ProTox-3 quickstart
Run your first ProTox-3 toxicity prediction.
You want a toxicity profile for a compound; this runs it in a few steps — enter a SMILES string, name, or drawn structure, choose which endpoints to predict, and read the results off a color-coded radar plot.
Not available in chat
ProTox-3 is deprecated on the platform and not registered with the orchestrator, so chat cannot call it. Use ADMET-AI for toxicity flags (hERG, Ames, hepatotoxicity) or DeepPurpose for specific endpoint predictions — existing ProTox-3 job results remain viewable.
Quick toxicity prediction with ProTox-3
Step-by-step workflow
Access the ProTox-3 tool
ProTox-3 no longer has a form in .bloom*'s Tools panel, and chat cannot call it (see the note above). The steps below describe the ProTox-3 workflow so you can read existing ProTox-3 results. ProTox-3 provides comprehensive toxicity profiling for small molecules.
Enter your compound
Option A: SMILES input
- Paste your canonical SMILES string
- Example:
CC(=O)Oc1ccccc1C(=O)O(aspirin) - .bloom* validates the structure automatically
Option B: Compound name
- Enter the PubChem compound name
- Example:
aspirin,caffeine,ibuprofen - System looks up the canonical structure
Option C: Draw structure
- Use the molecular sketcher
- Draw atoms, bonds, and functional groups
- Preview the structure before submission
Configure prediction settings
Endpoint selection:
- All endpoints (default): Complete 61-endpoint profile
- Acute toxicity only: LD50 and toxicity class
- Organ toxicity: Hepato/nephro/neuro/cardiotoxicity
- Custom selection: Choose specific endpoints
Submit and view results
Toxicity radar plot:
- Visual overview of all toxicity predictions
- Green = low risk, Yellow = moderate, Red = high risk
- Confidence scores shown for each endpoint
Detailed predictions:
- Per-endpoint toxicity classification
- Predicted values (e.g., LD50 in mg/kg)
- A confidence value per endpoint — an uncalibrated ordering hint, not an applicability-domain estimate
Similar compounds:
- Known toxic compounds with similar structures
- Toxicophore alerts and structural flags
- Reference compounds from training data
Export and integrate
Export formats:
- CSV with all predictions and confidence scores
- PDF report for documentation
- JSON for programmatic access
Integration:
- Send results to ADMET profiling workflows
- Combine with other safety predictions
- Include in compound decision reports
Example: Predicting caffeine toxicity
Shape of the result (illustrative — the values a run returns depend on the upstream model version, and an LD50 near 192 mg/kg falls in Class III, not Class IV, per the class table):
- Predicted LD50 in mg/kg, with its GHS class I–VI
- Hepatotoxicity: Active / Inactive, with a confidence value
- Mutagenicity: Active / Inactive, with a confidence value
- BBB permeability: Active / Inactive, with a confidence value
⚠︎ Read each endpoint on its own. Agreement with the bench differs sharply between them, the LD50 output is graded as a class and never as a number, and nothing here substitutes for the corresponding assay — see translatability.
Common use cases
Virtual screening triage
- Upload compound library (100-1000 compounds)
- Run all toxicity endpoints
- Filter for compounds with no high-risk flags
- Export prioritized list for further evaluation
Lead optimization
- Submit lead compound and analogs
- Compare toxicity profiles across series
- Identify toxicity-associated substructures
- Guide optimization to reduce toxicity
Safety screening
- Select compounds for preclinical development
- Generate comprehensive toxicity profiles
- Decide which assays to run first, and in what order
- Plan the confirmatory in vitro/in vivo studies — the predictions do not replace them, and are not a regulatory assessment
Troubleshooting
Invalid compound errors
- Check SMILES syntax with RDKit validation
- Ensure compound is a valid small molecule
- Remove salts and counterions
Low confidence predictions
- Compound may be outside model applicability domain
- Consider experimental validation for critical decisions
- Check similar compound coverage in training data
Read next
Was this page helpful?
How ProTox-3 results compare to the bench
What a ProTox-3 toxicity call has been shown to agree with, per endpoint, why its in vivo endpoints are not an in vivo comparison, and which endpoints have no published figure at all.
ProTox-3 troubleshooting
Diagnose invalid input, missing predictions, quota errors, and low confidence.