Note: This anchor describes the computational batch-integration method Harmony (Korsunsky et al., Nature Methods 2019), as applied to cell-profiling and image-based screening workflows. It is unrelated to the PerkinElmer Harmony imaging software.
The problem — Profiles acquired across plates, instruments, or labs carry systematic technical offsets that can dwarf the biological signal. In a Cell Painting screen spanning multiple acquisition days, a real two-fold phenotypic shift can be invisible beneath a four-fold plate effect. Naive mean-centring removes the offset but cannot account for batch-dependent scaling or rotational drift in high-dimensional feature space.
What it is / how it works — Harmony operates in an embedding space (typically the first principal components of the feature matrix). It assigns each sample a soft membership across a set of cluster centroids, then iteratively re-estimates those centroids after removing the contribution of batch labels. The correction is additive in embedding space: each sample is shifted toward a batch-integrated centroid position without distorting the local neighbourhood structure. The algorithm converges in a small number of iterations and scales to hundreds of thousands of wells.
In the Arevalo et al. (2024) benchmark of ten batch-correction methods on JUMP Cell Painting profiles across five technical-heterogeneity scenarios, Harmony and Seurat RPCA were the top two performers — a striking result given that both were originally designed for single-cell RNA-seq. Harmony's strength lies in handling moderate-to-high batch heterogeneity without collapsing phenotypic clusters; it consistently outperformed the domain-standard Sphering/whitening approach.
Where it fits in the pipeline — Harmony belongs at the correct stage, applied to PCA-reduced well-level profiles after feature normalisation and before similarity search, clustering, or Percent Replicating scoring. It returns corrected embedding coordinates; downstream tools that need feature-space vectors require projecting back — a step that should be explicit, not implicit.
Where it breaks — Harmony assumes that batch effects are distributed across clusters rather than being cluster-specific. When an entire biologically coherent group is present in only one batch (a common situation in targeted compound libraries), Harmony can conflate batch removal with phenotype removal. The Arevalo benchmark confirms that no correction method is biology-safe under high heterogeneity: validate the dial with a Percent Replicating check at each correction strength, and treat the maximally corrected result as the ceiling, not the default.