ProTox-3

ProTox-3 lets you predict toxicity across many endpoints — acute (LD50), organ toxicity, mutagenicity, carcinogenicity, and Tox21 pathways — from a compound's structure. Reach for it when toxicity is the whole question, for early safety screening or lead optimization. When you instead want toxicity folded into a broader absorption-and-metabolism profile, the general ADMET panels cover that; ProTox-3 is the dedicated toxicity predictor. See what you can achieve.

ProTox-3 integration in .bloom* platform

ProTox 3.0 is a comprehensive web server for the prediction of toxicity of chemicals, developed at the Charité - University Medicine Berlin. Published in Nucleic Acids Research (July 2024), it incorporates molecular similarity and machine learning models for predicting 61 toxicity endpoints.

Citations

Banerjee P, Kemmler E, Dunkel M, Preissner R. ProTox 3.0: a webserver for the prediction of toxicity of chemicals. Nucleic Acids Research. 2024 Jul 5;52(W1):W513-W520. doi: 10.1093/nar/gkae303

Key capabilities

ProTox-3 provides predictions across multiple toxicity categories:

Acute toxicity

• Oral toxicity (LD50): Lethal dose prediction with toxicity class assignment
• Toxicity classes: Classification into 6 toxicity classes (I-VI)

Organ toxicity

• Hepatotoxicity: Liver damage potential
• Neurotoxicity: Nervous system effects
• Nephrotoxicity: Kidney toxicity risk
• Cardiotoxicity: Cardiovascular safety
• Respiratory toxicity: Lung damage potential

Toxicological endpoints

• Mutagenicity: Ames test prediction
• Carcinogenicity: Cancer-causing potential
• Cytotoxicity: Cell death induction
• Immunotoxicity: Immune system effects
• BBB permeability: Blood-brain barrier penetration
• Ecotoxicity: Environmental toxicity
• Clinical toxicity: FDA-approved drug toxicity
• Nutritional toxicity: Dietary compound safety

Toxicological pathways (12 AOPs)

• Adverse outcome pathway predictions for mechanistic understanding
• Tox21 assay endpoint predictions

Molecular targets

• 15 toxicity targets: Known toxicity-related protein targets
• 14 molecular initiating events (MIEs): Initial molecular interactions
• 6 metabolism targets: Drug metabolism enzyme interactions

Webapp integration

Input methods in .bloom*

Direct compound entry:

• SMILES string input with validation
• PubChem compound name lookup
• ChemDoodle molecular sketcher for drawing structures

Batch processing:

• CSV file upload with multiple compounds
• SDF file support for 3D structures
• Integration with other .bloom* chemistry tools

Prediction outputs

Toxicity radar plot:

• Visual summary of all toxicity predictions
• Color-coded risk levels (green/yellow/red)
• Confidence scores for each endpoint

Network visualization:

• Toxicity target network plot
• Pathway interaction visualization
• Mechanistic toxicity insights

Detailed reports:

• Per-endpoint predictions with confidence scores
• Similar known toxic compound identification
• Structural alerts and toxicophore detection

When to use ProTox-3

Drug discovery applications

Early-stage safety screening:

• Filter compound libraries for toxicity liabilities
• Prioritize compounds with favorable safety profiles
• Guide medicinal chemistry optimization

Lead optimization:

• Assess toxicity impact of structural modifications
• Balance potency improvements with safety
• Identify toxicity-associated substructures

Hazard screening

Safety assessment:

• Flag organ-specific toxicity liabilities for experimental follow-up
• Prioritize which assays to schedule first
• Generate structured hypotheses for a weight-of-evidence discussion

Environmental assessment:

• Ecotoxicity predictions for environmental impact
• Chemical safety evaluation for industrial compounds

Performance characteristics

Model validation

• The 2024 ProTox-3 paper reports cross-validation and external-validation figures for its 28 newly added models. The other 33 — including acute toxicity, hepatotoxicity, mutagenicity, carcinogenicity, cytotoxicity, immunotoxicity and the 12 Tox21 pathways — are carried over, and their published figures come from ProTox-II (2018) and ProTox (2014)
• Every published figure is author-reported, with no sample size given in primary text, and agreement varies widely by endpoint — from AUC-ROC 0.95 for cardiotoxicity down to 0.74 with κ = 0.29 for immunotoxicity
• Per-endpoint figures with their version, metric and caveats: translatability

Confidence scoring

• Each endpoint carries a confidence or probability value from the upstream server
• The three most structurally similar compounds from the training data are returned alongside the acute-toxicity call
• ⚠︎ No published study calibrates the confidence score or relates it to applicability-domain distance. Read it as an ordering hint within one endpoint, not as a probability and not as an applicability-domain estimate

Integration with .bloom* ecosystem

Upstream data sources

• DeepPurpose: Evaluate toxicity of generated compounds
• REINVENT: Safety assessment of de novo designs
• ZINC/PubChem: Toxicity screening of database compounds

Downstream applications

• ADMET-AI/ADMETlab 3.0: Comprehensive ADMET profiling
• AiZynthFinder: Synthetic route safety assessment
• Decision support: triage and assay scheduling, never a go/no-go on a prediction alone

Resources

• Web server: https://tox.charite.de
• Publication: Nucleic Acids Research 2024
• Free access: No login required

Chat Usage

Example Prompts

• "Predict toxicity for aspirin and ibuprofen using ProTox3"
• "Run ProTox3 with organ toxicity models on my ADMET hits"
• "Check CYP inhibition and hepatotoxicity for these 10 compounds"

Workflow Chaining

• Feeds into: Safety Passport (contributes toxicity data), Compound Passport
• Accepts from: PubChem (compound names/SMILES), ADMET screening results, REINVENT/GenMol (generated molecules)

FAQ

Read next