The problem. Reference atlases keep growing, and every new dataset traditionally means re-running integration over the whole combined corpus — expensive, and it needs the original data on hand, which is often impractical or restricted. You want to map a fresh query onto an existing atlas without retraining from scratch or sharing the reference’s raw cells.

The idea. scArches — single-cell architectural surgery — treats a pretrained generative reference model (like scVI/totalVI) as fixed and adds a few new weights that adapt it to the query, training only those. This “surgery” learns the query’s batch-specific variation while preserving the reference’s biological structure, using orders of magnitude fewer parameters than de novo integration. The query lands in the reference’s coordinate system, ready for label transfer, and only the model — not the reference data — needs to travel.

Why it matters. This is transfer learning arriving in single-cell analysis, and it fits the reference-mapping thread that runs from SingleR and Seurat v4 through scANVI — but reframed for a world of shared, ever-growing atlases. The data-privacy angle (map without redistributing raw cells) matters for clinical settings like the STU’s. It’s also a fitting close to two days of integration: the point where atlases become reusable infrastructure.

Verdict. A widely-used, efficient reference-mapping framework; it assumes a good pretrained reference exists and that the query isn’t wildly outside its scope. Read it as fine-tuning an atlas to meet new data — integration turned into a reusable model.