The problem — Plate effects, scanner-to-scanner variation, batch-to-batch reagent drift, and well-position confounders all land in the feature table alongside biology. Two compounds measured on different plates can look more different from each other than two unrelated compounds measured on the same plate — not because of their mechanism, but because of when they were imaged. Without correction, cross-plate or cross-site analysis is confounded by construction.
What it is / how it works — Batch correction estimates and removes the systematic component of technical variation. The toolbox spans methods first developed for single-cell RNA-seq and adapted to image-based profiling: sphering (whitening the feature covariance to remove per-plate structure), Harmony (iterative cluster-aware embedding correction; Korsunsky et al.), pyComBat (linear mixed-model removal), and scVI (deep variational for spatial omics). Arevalo et al. benchmark ten methods on JUMP-scale Cell Painting data across five scenarios — single-lab temporal batches through multi-lab multi-instrument — and find Harmony and Seurat RPCA best, but critically: no method removes batch without some loss of biology. Correction is a dial, not a switch.
Where it breaks — The symmetric failure is over-correction: a method parameterized too aggressively flattens the phenotypic distances that constitute the readout, producing a profile set where replicates agree but perturbations are indistinguishable. The validation gate is explicit: Percent Replicating must increase after correction (or at least not decrease) — the correction that maximizes reproducible signal is correct, not the one that maximizes mixing. See Batch Correction Without Erasing Biology for the operating stance. For spatial omics and digital pathology, the equivalent failure is domain adaptation that harmonizes a scanner signature at the cost of tissue-type separability.