ProTox-3 outputs & interpretation
Understanding toxicity predictions and confidence scores.
ProTox-3 returns per-endpoint toxicity classifications, an LD50 class, and confidence scores — shown as a radar plot and interaction network, exportable as CSV. Use these to triage a screening library, compare analogs during lead optimization, or flag structural alerts before committing to experimental work.
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.
Output formats
Toxicity radar plot:
- Visual summary across all 61 endpoints
- Color-coded risk levels
- Interactive hover for details
Network visualization:
- Toxicity target interaction networks
- Pathway connections and mechanisms
- Structural alert highlighting
CSV export:
- All predictions with confidence scores
- Endpoint names and classifications
- Similar compound information
Interpreting toxicity classes
Acute oral toxicity (LD50 classes):
| Class | LD50 (mg/kg) | Description |
|---|---|---|
| I | ≤5 | Fatal if swallowed |
| II | 5-50 | Fatal if swallowed |
| III | 50-300 | Toxic if swallowed |
| IV | 300-2000 | Harmful if swallowed |
| V | 2000-5000 | May be harmful if swallowed |
| VI | >5000 | Non-toxic |
Binary endpoint predictions:
- Active/Inactive: Classification result
- Probability: 0-1 scale likelihood
- Confidence: Model reliability estimate
Confidence interpretation
⚠︎ The confidence value is not a calibrated probability. No published study relates it to applicability-domain distance or measures whether high-confidence calls are more often correct, so the bands below are a reading convention and not a validated scale. Use the value to order compounds within a single endpoint; do not compare it across endpoints, and do not treat a high value as licence to act without an assay. See translatability.
High relative confidence (>0.8):
- The input sits closer to the chemistry behind that endpoint's model
- Worth looking at first, and worth scheduling the corresponding assay earlier
Medium relative confidence (0.5–0.8):
- Use as supporting information alongside the similar-compound list and structural alerts
- Experimental confirmation still required before any decision
Low relative confidence (<0.5):
- Likely novel chemical space for that model
- Read the call as absent rather than negative
Structural alerts
Toxicophore detection:
- Known toxicity-associated substructures
- Reactive functional groups
- Metabolic activation sites
Similar compound analysis:
- Top similar compounds from training data
- Known toxicity of similar structures
- Structure-activity insights
Canvas Visualization
When a ProTox3 job completes, a result card appears on the canvas.
- Card size: 350×300px
- Border color: Orange (#FF9800)
- Visualization: Toxicity summary with class indicator
Tier 2 Metrics
- Compound count
- Toxicity class (I–VI)
- Predicted LD50 (mg/kg)
- Confidence percentage
Toxicity Class Interpretation
| Class | Label | LD50 Range |
|---|---|---|
| I | Fatal | ≤ 5 mg/kg |
| II | Fatal | 5–50 mg/kg |
| III | Toxic | 50–300 mg/kg |
| IV | Harmful | 300–2,000 mg/kg |
| V | May be harmful | 2,000–5,000 mg/kg |
| VI | Non-toxic | > 5,000 mg/kg |
Read next
Was this page helpful?
ProTox-3 limits & performance
Batch sizes, daily quota, throughput, and endpoint-driven cost.
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.