Task-Aware Qualification of Cell Painting Feature Spaces for Macrophage Phenotypic Assay Development

This dataset accompanies the study “Task-Aware Qualification of Cell Painting Feature Spaces for Macrophage Phenotypic Assay Development”. It contains the data and analysis outputs used to qualify zero-shot Cell Painting feature spaces for bidirectional M2-to-M1 and M1-to-M2 macrophage reprogramming screening. The repository includes postprocessed model-derived feature profiles from three independent experimental batches, nuclei-count quality-control data, harmonized analysis-ready profiles, retrieval-analysis outputs, statistical result tables, reference-well-count and field-of-view robustness analyses, figure source data and publication-ready figures, together with the Python scripts and software-environment specifications required to reproduce the computational workflow. Four pretrained image-based models—CellPaintSSL, OpenPhenom, uniDINO, and DINOv2—were evaluated in combination with two well-level aggregation methods and ten postprocessing workflows. The qualification framework evaluates preservation of the cross-batch M1/M2 biological contrast together with residual batch-label retrieval and subsequently assesses robustness to reference-control allocation and image-acquisition depth. The final operational configuration identified for the predefined screening task uses DINOv2 with median well-level aggregation and MAD_TVN postprocessing, 8 reference wells per macrophage state and experimental batch, and 6 fields of view per well. The deposited repository also retains supporting results for the alternative Pareto-selected feature space and evaluated reference-well and field-of-view designs. This version additionally includes MPH_STAGE06_PAIRWISE_PUBLIC.tar.gz, containing the current-canonical post-selection pairwise batch-composition robustness analysis for the Stage04-selected DINOv2 median N9 MAD_TVN and N10 ZCA_INT feature spaces.