JUMP Cell Painting chemical dataset: Morphological map of 115,796 compounds in human cells

Key results

1. 01

   ShownPhenotypic signal persists across 12 laboratories, though cross-laboratory replicates detect fewer active reference compounds than within-laboratory replicates; activity is only partly explained by cell loss.

   17,069 of 112,462 compounds (15.2%) separated from DMSO at FDR < 0.10 (copairs mAP). On the 301-compound Target-2 plate, 77% were active with within-laboratory replicates vs 51% with cross-laboratory replicates after Harmony correction (a 34% reduction). Active compounds had median 411 vs 457 cells/mm^2; cell density vs activity Spearman rho = -0.33. · Phenotypic signal is detectable and reproducible…; Phenotypic activity and consistency metrics; Within-source and pooled-source reproducibility…

2. 02

   ShownCompounds sharing a mechanism of action (MOA) or target tend to have similar profiles, supporting guilt-by-association mechanism hypotheses; using data across laboratories needs more replicates.

   For MOA-annotated active compounds, the nearest phenotypic neighbor shared the MOA in 21.2% of cases vs 1.3% under label permutation (16-fold), with same-scaffold compounds excluded. Many targets, MOAs and disease areas were consistent across nine annotation sets. 2-3 replicates within a laboratory, or 4-5 across laboratories, gave 90% of maximal MOA consistency. · Image-based profiles can predict the targets and…; Phenotypic activity and consistency metrics; Within-source and pooled-source reproducibility…; Structure-phenotype analysis

3. 03

   SuggestedA small separate experiment queried against JUMP linked two PPARG inverse agonist series to cell-cycle targets; follow-up assays indicated G1 arrest and senescence resembling the CDK4/6 inhibitor palbociclib.

   Five BAY compounds profiled in duplicate in UM-UC-9 cells best matched JUMP compounds annotated to CDK, RRM and TOP2A; PPARG-targeting compounds were mostly filtered out as inactive in U2OS. UM-UC-9 queries matched better than U2OS ones. Flow cytometry showed more G1 and fewer S/M cells; p-Rb (S780) fell and SASP genes (TNF, CXCL10, CCL5, IL6) and beta-gal-positive cells rose. · Image profile matching identifies PPARG inverse…; PPARG plates Pearson correlation; PPARG plates target association; Cell Cycle Analysis

4. 04

   ShownScaffolds partly predict morphology, but per-compound structural and morphological neighborhoods overlap little; the data yield lists of on/off activity cliffs and phenotypic cliffs, the latter enriched for pairs without a shared target.

   158/262 (60.3%) Bemis-Murcko scaffolds were more self-similar than size-matched random groups. Per compound, k=50 neighborhood Jaccard overlap was 0.010 vs 0.0015 expected. 19% of structurally similar pairs were on/off activity cliffs; 4,089 of 12,920 similar active pairs had profile cosine < 0.3, and 59.7% of 159 annotated cliff pairs shared no target vs 29.5% of concordant pairs (OR 3.5). · Image-based profiles reflect structure-activity…; Structure-phenotype analysis; Activity cliff detection

5. 05

   ShownMorphology adds little over chemical structure for predicting pan-assay interference (PAINS) status, organism-level toxicity (DICT, DILI), pharmacokinetics or most ToxCast and ChEMBL assay activities; mitotoxicity and a few cell-based nuclear-receptor and stress assays are exceptions.

   PAINS (scaffold split): ROC-AUC 0.634 from profiles vs 0.634 from MW/LogP/TPSA and 0.547 from cell count (0.660 combined). 95 ToxCast assays: median AUC 0.611 vs 0.674 for fingerprints and 0.538 for cell count; 51 ChEMBL targets: 0.563 vs 0.584. ECFP4 beat profiles for cardiotoxicity (DICT), liver injury (DILI) and pharmacokinetics. Exceptions: CAR 0.854 vs 0.607, FXR 0.807 vs 0.687, mitotoxicity. · Image-based profiles predict only certain…; Predicting Activities from Cell Painting…; PAINS prediction