ProTox-3 troubleshooting

This page helps you diagnose why a ProTox-3 job errored, returned no predictions, or produced low-confidence scores — most often traceable to SMILES syntax, a failed compound-name lookup, the daily query quota, or compounds in novel chemical space outside the model's applicability domain.

Input and validation failures

Invalid SMILES:

• Validate SMILES syntax before submission and prefer canonical SMILES for consistency.
• Check for valid atom types and correct valence; ProTox-3 standardizes structures, but malformed strings fail during processing.
• Invalid SMILES are reported per compound, so a single bad row does not fail the whole batch — check which compound the log flags.

Compound name did not resolve:

• Name entry relies on a PubChem lookup; uncommon or misspelled names can fail to resolve.
• Switch to a SMILES string, or verify the exact synonym in PubChem first.

No model selected:

• At least one toxicity model must be selected in addition to at least one compound; the form blocks submission otherwise. See inputs & parameters.

No predictions returned

• The compound may fall outside the model's applicability domain — very large or unusual structures have limited coverage.
• Confirm the structure is a valid small molecule; peptides and biologics are not supported.
• Re-run a single known compound (for example, aspirin) to confirm the pipeline itself is healthy before debugging the batch.

Quota and rate limits

• ProTox-3 enforces a shared 250 queries per day limit; a large comprehensive run can exhaust it quickly.
• The webapp warns at 45 compounds precisely to flag quota consumption before you submit.
• If a job stalls or errors after heavy use, you may have hit the daily cap — wait for the reset rather than resubmitting. See limits & performance.

Low confidence scores

• Novel chemical space may have limited training coverage, lowering confidence.
• The confidence value is uncalibrated and no published study ties it to applicability-domain distance, so treat a low value as "no usable call for this endpoint" rather than as a negative result. See translatability.
• Compare against similar known compounds and the reported structural alerts before making a call — and confirm with the corresponding assay, which no confidence level removes the need for.

If a job fails

1. Open the job log and read the error category (validation vs. service).
2. Validate the input SMILES externally to rule out syntax issues.
3. Re-run with a smaller batch or a focused endpoint set to isolate the failing step.
4. If failures persist, contact support with the job ID and the first few input rows.

Quality checks

• Compare predictions across a compound series to spot outliers.
• Validate against known toxic and non-toxic reference compounds.
• Review structural alerts for mechanistic insight into flagged endpoints.
• Weigh confidence scores when acting on individual predictions.

Read next