The problem. Combining single-cell datasets (Seurat’s anchors, day 10, solve this too) means removing batch effects without erasing biology. Anchor- and MNN-based methods work but can be heavy on large datasets, and you often want to integrate after dimensionality reduction, iterating quickly rather than reprocessing from raw counts.

The idea. Harmony operates in PCA space. It iteratively clusters cells in the embedding, computes batch-specific correction factors that pull cells of the same soft cluster together across batches, and repeats until batches are mixed but clusters preserved. Because it works on the low-dimensional embedding rather than the full matrix, it’s fast and scales to large atlases — while a soft-clustering penalty guards against over-correcting away real differences.

Why it matters. Harmony is one of the most-used integration tools and a standard Scanpy/Seurat step, so it’s a default I should understand rather than invoke blindly. It’s the fast, embedding-space counterpart to Seurat’s anchors and Scanorama’s stitching — three answers to the same batch problem, differing in where and how aggressively they correct. Knowing the trade-offs is knowing when to trust the integrated space.

Verdict. Foundational for practical single-cell integration; its speed made routine integration feasible. As with all batch correction, over-integration is the risk to watch. Read it as the efficient, widely-defaulted member of the integration toolkit.