The problem. Combine single-cell datasets from different runs, labs, or technologies and batch effects dominate: cells cluster by experiment, not biology. Earlier correction methods assumed the batches shared the same population composition — a bad assumption when one sample is enriched for a cell type the other lacks. You need a correction that works even when the mixtures differ.

The idea. Two cells form a mutual nearest-neighbour (MNN) pair if each is among the other’s closest neighbours across batches — a signal that they’re the same underlying cell type. The differences across many such pairs estimate the batch-effect vector, which is then subtracted to align the datasets. Because MNNs are found locally, the method tolerates different population compositions between batches, unlike global approaches.

Why it matters. MNN is the conceptual ancestor of the anchor- and neighbour-based integration that runs through the single-cell reading — Seurat’s anchors, Scanorama, and the batch correction inside Harmony all echo this mutual-neighbour logic. Integration is the unglamorous step that decides whether a multi-sample atlas means anything, and for pipeline-minded work it’s the QC gate before downstream analysis. It’s where “combine these datasets” became a solved-enough problem.

Verdict. A landmark that reframed batch correction and seeded a family of methods; the correction assumes at least some shared cell types exist to anchor on. Read it as the mutual-neighbour insight the whole integration toolkit was built on.