GenMol troubleshooting

This page helps you diagnose the most common GenMol problems — invalid SMILES, low validity or diversity, over-aggressive quality filtering, job timeouts, unexpected scaffold decorations, and SAFE fragment ingestion failures — and points to the parameters to adjust for each.

Common issues and fixes

GenMol uses masked discrete diffusion on SAFE molecular representations. Understanding common issues helps you configure jobs correctly and interpret results appropriately.

Invalid SMILES errors

• Verify your SMILES string parses correctly using the real-time validator in .bloom* or an external tool like RDKit
• Check for unmatched parentheses, invalid ring closure digits, or incorrect element symbols
• Ensure stereochemistry notation (@, @@, /, \) is syntactically correct
• Remove whitespace, newlines, or extraneous characters from the SMILES string
• For scaffold decoration, confirm SAFE attachment points use valid numbered bracket syntax (e.g., [1])

Low validity rate

• If fewer than 80% of generated molecules are valid, try reducing the temperature parameter (e.g., from 1.0 to 0.7) for more conservative sampling
• Increase the number of diffusion steps (num_steps) for more refined denoising
• In fragment-constrained mode, ensure the fixed fragments are chemically reasonable and have valid attachment points
• For linker design, verify both fragment SMILES are individually valid before submission

Low diversity in generated molecules

• Increase the temperature parameter (e.g., from 1.0 to 1.3) for more diverse sampling
• In lead optimization mode, reduce the similarity threshold to allow more structural variation
• Increase the number of molecules generated; larger batches naturally produce more diverse sets
• If using PMO mode, the oracle may be driving convergence to a narrow chemical space; this is expected behavior for goal-directed generation

Fewer usable molecules than expected

The quality filters remove molecules from the returned set, so a low count can come from the filters as well as from the generation. enforce_validity and enforce_uniqueness are on by default; apply_druglikeness_filter, min_qed and max_sa_score bind only when you set them, except in lead_optimization, which applies its own QED floor and SA ceiling.

• Read the quality_filters block in the result: it reports how many molecules went in, how many came out, and how many each filter removed
• If a lead run returns nothing, widen min_qed — the QED floor is the bound most likely to have emptied the set
• If your own QED cutoff is removing most molecules, lower the temperature for more conservative sampling, then re-apply the cutoff
• Generate more molecules to compensate for your own filtering loss
• For makeability, run AiZynthFinder on the survivors rather than relying on a score — see how GenMol results compare to the bench

Job timeouts or long runtimes

• Reduce the number of molecules per job (try 500 instead of 5,000)
• Lower the number of diffusion steps (num_steps) to reduce per-molecule generation time
• For very large campaigns, split into multiple jobs of 1,000 molecules each
• Check cluster capacity and retry during lower-traffic windows if queue times are long

Scaffold decoration produces unexpected R-groups

• Verify that SAFE attachment points are correctly placed on the scaffold
• Check that the scaffold SMILES is in canonical form
• Ensure attachment point numbering is consistent and sequential
• Try generating with a lower temperature to produce more conservative decorations
• Review the SAFE encoding of your scaffold to confirm fragment boundaries are where you expect them

PMO mode converges to unreasonable molecules

• Some oracles reward extreme property values. This is reward exploitation: when a computed oracle is the only judge, the optimizer finds that oracle's blind spots, and a high oracle score becomes a statement about the oracle
• Limit the optimization budget to prevent excessive optimization into unrealistic chemical space
• Set min_qed on the run, or apply your own post-hoc cutoffs (for example QED > 0.3) to the returned properties
• Inspect the 2D structures. Spot-checking is the cheapest control available and catches failures no metric reports

SAFE fragment ingestion failures

• Fragments from generated molecules are automatically ingested into the project fragment library
• If fragments are not appearing, check that the job completed successfully (not just partially)
• Verify the FEATURE_SAFE_INGESTION flag is not disabled in your environment
• Very small fragments (fewer than 3 heavy atoms) may be filtered during ingestion

Still stuck?

• Provide the job ID, generation mode, SMILES input (if applicable), and parameter settings
• Note whether the issue is at submission, during generation, or in the output
• Share the validity rate and any error messages from the job log
• Ask Ivy for real-time guidance on parameter tuning

Read next