ADMETlab 3.0 troubleshooting

This page helps you diagnose the issues that most often interrupt an ADMETlab 3.0 run — invalid SMILES, oversized batches, slow queues, missing or partial scores from out-of-domain compounds — and how to interpret its toxicity predictions.

Common errors

ADMETlab 3.0 uses a multi-task DMPNN architecture trained on 400,000+ compounds. Most errors relate to input format or compounds outside the model's training domain.

• Invalid SMILES: See the rejected rows in the CSV; regenerate canonical SMILES and retry. Ensure proper valence states and bond orders.
• Too many molecules: Reduce batch size to within workspace limits (recommended: 50-100 compounds); oversize uploads are blocked before queueing.
• Missing scores: Check for out_of_domain or standardization_failed flags; rerun with preprocessing enabled. Out-of-domain compounds fall outside the training data chemical space.
• Slow queue: Split into multiple jobs and avoid peak hours; monitor queue depth on the job page. The extended 119-endpoint panel takes longer than core endpoints.
• Partial results: Some endpoints may return null for specific compounds if applicability domain checks fail. This is expected for molecules dissimilar to training data.

When a job fails

1. Open the job log to confirm failure category and model version.
2. Rerun with a smaller batch (50-100 compounds) or fewer endpoints to isolate the issue.
3. Check if specific compounds cause failures by testing subsets.
4. Verify SMILES validity using external validation tools (e.g., RDKit).
5. If still failing, share the job ID plus first 10 rows with support for reproduction.

Prevent repeat failures

• Keep a small regression set (20-50 molecules) to validate parameter tweaks.
• Save job notes describing why you reran or changed parameters.
• Compare canonical SMILES between runs to spot preprocessing differences.
• Document which endpoint bundle (core vs. extended 119 endpoints) was used for reproducibility.

Interpreting toxicity predictions

The 31 new toxicity endpoints in v3.0 require careful interpretation:

• Nephrotoxicity, neurotoxicity, ototoxicity, hematotoxicity: These organ-specific predictions indicate potential liability. Higher scores suggest greater risk.
• Applicability domain: Toxicity predictions are most reliable for drug-like molecules similar to the training data.
• Confidence scores: Use low-confidence predictions as flags for experimental follow-up rather than definitive assessments.
• Cross-validation: Consider comparing predictions with ADMET-AI or other tools for critical safety decisions.

Read next