The problem — A raw CellProfiler feature table contains hundreds of measurements spanning orders of magnitude: intensity values in the thousands, texture scores near one, shape eccentricities between zero and one. Without normalizing, downstream distance metrics and dimensionality reduction are dominated by the features with the largest absolute range. Plate-to-plate reagent variability and per-acquisition intensity drift land in the same table, indistinguishable from biology. Two plates run a week apart will look more different than two biologically distinct perturbations run the same day — not because of mechanism, but because of when the assay was run.
What it is / how it works — Normalization standardizes each feature independently, per plate. The field-standard method is robust z-scoring against negative controls: for each feature, subtract the median and divide by the MAD (median absolute deviation) of DMSO wells on that plate. pycytominer exposes this as RobustMAD — the formula is (x − median) / MAD — with a small epsilon to prevent division by zero on near-constant features. Normalizing against DMSO anchors each plate's coordinate system to its own untreated population rather than to global statistics that vary across batches.
The choice of reference population is a scientific decision, not a default:
- DMSO wells only — removes plate-wide drift while preserving perturbation signal relative to the untreated baseline. Preferred when the question is "how does compound X perturb cells?"
- Whole-plate normalization — uses all wells, including treated, as the reference. Appropriate when DMSO coverage is sparse but introduces subtle signal compression if strong perturbations shift the plate distribution.
Both options are available in pycytominer; document which was used, because downstream Percent Replicating scores and batch correction results are not comparable across the two.
Where it breaks — Normalization inherits every upstream failure silently. An uncorrected illumination gradient, a contaminated DMSO row, or a segmentation collapse in the negative-control wells all propagate into the normalized feature table with no error raised. The scale disappears; the artifact does not. Always validate that DMSO replicates cluster tightly in normalized space before proceeding to batch correction.