Cell Painting: hits that were well positions
Most 'active' compounds sat in outer wells, and replicate compounds failed to find each other. The phenotypes were mostly position effects.
- Client
- Drug discovery biotech
- Data
- Cell Painting, 2,000 compounds, 24 plates, 384-well
- Timeline
- 4 weeks
- Outcome
- Hit list rebuilt; 2 MoA clusters advanced
Anonymized composite case. Details are combined and altered to protect client confidentiality.
The problem
The team's in-house analysis normalized features to the whole plate and called 300 hits. Medicinal chemists could not find any structure-activity relationship, and repeat plates gave different hits.
What we found
- Hits were enriched about fourfold in edge wells. Evaporation changed cell density and morphology.
- Compound layout was not randomized across plates, so position and compound identity were linked.
- Replicate retrieval was near random (mAP close to baseline) for most hits, which meant the profiles were not reproducible.
How we rescued it
- Normalized to on-plate DMSO controls, corrected well-position effects, and sphered profiles across plates.
- Selected features with pycytominer and removed density-confounded features.
- Scored phenotypic activity and replicate mAP with permutation testing, and kept only reproducible compounds.
- Mapped mechanism-of-action similarity against annotated reference compounds.
The result
The reproducible hit list shrank to 64 compounds that formed coherent clusters. Two clusters matched tubulin and HDAC reference phenotypes and lined up with chemistry series. We also designed a randomized plate layout for the next screen.
| Measure | Result |
|---|---|
| Hits before | 300 |
| Reproducible hits after | 64 |
| Edge-well enrichment before | ~4x |
Have a project like this?
We audit failed or questioned analyses, explain what went wrong, and rebuild them reproducibly, often within days of a reviewer deadline.
- Service
- Cell Painting and single-cell phenomics
- Price
- From $79/h