The problem. Most integration aims to erase differences between datasets so cells line up. But sometimes the differences are the biology — a disease sample versus control, a different species, RNA versus a chromatin readout. You want a method that finds common structure to align on while explicitly preserving, not discarding, what’s unique to each dataset.

The idea. LIGER uses integrative non-negative matrix factorization (iNMF): it factorizes multiple datasets into a set of shared metagene factors plus dataset-specific factors. Cells are then matched in the shared factor space (via a neighbour-graph step), giving joint clusters, while the dataset-specific factors quantify what each sample contributes on its own. Applied to brain data, it aligns cell types across modalities and highlights modality-specific signals.

Why it matters. This is the multi-view sibling of the factor and integration methods around it — it shares MOFA+’s decompose-into-factors instinct and the alignment goal of Harmony and Scanorama, but its shared-versus-specific split is the distinctive move. For comparative studies, and for the STU contrasting spatial modalities or conditions, keeping the differences interpretable is often the whole point.

Verdict. A widely-used integration method with a genuinely different philosophy; NMF factor counts and regularization need tuning, and interpretation takes care. Read it as integration that aligns without flattening — shared and specific held apart on purpose.