RXNFP troubleshooting

This page helps you diagnose reaction SMILES parsing failures, unexpectedly similar fingerprints, and empty similarity-search results.

RXNFP troubleshooting

Invalid reaction SMILES

Symptom: a reaction in your batch has is_valid: false with a parsing error.

Fixes:

• Ensure the reaction contains >> separating reactants from products.
• Validate each component as a valid SMILES string.
• Remove whitespace and newlines from the SMILES.
• Check for balanced parentheses and ring closures.
• A failed reaction doesn't fail the whole batch — the other reactions in the same call still return fingerprints; fix or drop the failed row and resubmit if you need it.

Fingerprints look too similar

Symptom: cosine similarity between reactions you'd expect to be unrelated comes back high.

Causes:

• Very short SMILES: simple reactions with few atoms produce sparse transformer representations that overlap more than expected.
• Same functional groups: reactions sharing major functional groups will have partial fingerprint overlap even when the transformation differs.
• There's no alternative model to switch to — the fingerprint is always generated by the same fixed RXNBERT model, so a similarity result you don't trust needs to be checked against Reaction Classifier's named classification instead of a different RXNFP setting.

Similarity search returns no results

Symptom: similar_reactions comes back empty.

Fixes:

• Lower min_similarity — the default (0.7) can exclude everything for a genuinely novel transformation.
• Raise top_k if you suspect matches exist just below the current cutoff.
• Remember similarity search only matches against the platform's reference index (built from ORDerly reaction data); a transformation with no close analog there will legitimately return few or no matches.

Query reaction fails to fingerprint

Symptom: the similarity-search job fails before returning any matches.

Fix: the query reaction SMILES has the same validity requirements as fingerprinting — confirm it parses on its own with a plain fingerprint call first, then retry the similarity search.

Getting help

If issues persist:

1. Check the job logs for the specific error message.
2. Validate your reaction SMILES format.
3. Describe the issue to Ivy.

Read next