The Correction That Erased the Signal
Single-Cell · Data Integration
The Correction That Erased the Signal
Batch correction is supposed to remove the artifact and leave the biology. Push it too hard and it removes both — and the plot still looks beautiful.
You ran three 10x lanes over two weeks. The first two shared a reagent lot; the third came after a chemistry upgrade. On the raw UMAP the cells split cleanly by lane before they split by anything you care about — a textbook batch effect. So you reach for an integration method, rerun, and the lanes dissolve into one continuous manifold. Clean. Publishable.
The uncomfortable question is whether the method removed a technical artifact or quietly homogenized real biology along with it. Both look identical on a UMAP: cells that used to sit apart now sit together. The plot cannot tell you which kind of separation you just deleted — and that is the entire problem with judging integration by eye.
Two objectives pulling in opposite directions
Why this is a trade-off, not a setting
Every integration method optimizes two goals that are fundamentally in tension. The first is batch removal: make cells mix across batches so lane, lot, and platform stop driving the structure. The second is biological conservation: keep genuinely distinct cell states, subtypes, and trajectories from collapsing into each other.
These pull against one another because both express themselves as "cells being close in the embedding." A method has no innate way to know whether two nearby cells are close because they are the same cell type measured on two platforms (good — merge them) or because they are two related states the correction over-smoothed (bad — you just lost a subtype). Turn the batch-removal dial up and you win the first objective while bleeding the second.
This is why "did integration work?" is not answerable from a single UMAP, and why the field settled on paired metrics that score the two axes separately.
Measuring both sides
The metric vocabulary you actually need
Rather than one score, robust evaluation reports a batch-mixing metric and a biology-preservation metric side by side. The most widely used are worth knowing by name, because collaborators and reviewers will cite them:
| Metric | Measures | Reads well when… |
|---|---|---|
| kBET | Batch mixing in local neighborhoods | Each cell's neighbors match the global batch proportions |
| iLISI | Batch diversity around a cell | Neighborhoods contain multiple batches |
| cLISI | Cell-type purity around a cell | Neighborhoods stay single-cell-type |
| ASW (cell type) | Separation between labeled types | Types remain well-separated after integration |
| ARI / NMI | Cluster–label agreement | Post-integration clusters still recover known labels |
The pattern is the point: iLISI and cLISI are near-mirror images computed the same way — one rewards mixing across batches, the other rewards not mixing across cell types.[4] A method that aces one and fails the other has told you exactly which objective it sacrificed. The large atlas-integration benchmark from the Theis group formalized this by combining fourteen such metrics into a weighted batch-removal and bio-conservation score, precisely so that no method could win by over-correcting.[1]
What the benchmarks actually found
A finding that should change your defaults
Two results recur across independent benchmarks and are worth carrying into your own pipeline decisions.
First, on the mechanics of preprocessing: selecting highly variable genes before integration tends to help, while aggressive scaling pushes methods to prioritize batch removal over conservation of biological variation.[1] That is not a quirk — it is the trade-off surfacing as a preprocessing choice. The knob you thought was cosmetic moves you along the exact axis that matters.
Second, on method choice: no single tool wins everywhere, and difficulty of the task changes the ranking. On simpler integration tasks, linear-embedding methods like Harmony and Seurat perform strongly and are fast.[2] On complex tasks with strong, nested batch effects, deep-learning methods — scVI, and its label-aware extension scANVI, alongside Scanorama and scGen — tend to hold biology together better. The label-aware methods, which get to see cell-type annotations during integration, are the ones that most reliably integrate across strong batch effects without dissolving subtypes.[1]
A working default, not a verdict
Start with Harmony. It is fast, scales, and repeatedly lands among the top performers on straightforward tasks[3] — a sensible first pass before you spend GPU time.
Escalate to scANVI or scVI when the task is hard — many donors, mixed platforms, single-cell plus single-nucleus — or when a first pass visibly smears subtypes you expected to stay distinct. If you have trustworthy cell-type labels, a label-aware method is usually worth it.
Never accept the result on the UMAP alone. Report a batch metric and a bio-conservation metric together, every time.
The failure mode to watch
When a clean plot is the warning sign
The dangerous outcome is not the messy integration — that one announces itself. It is the over-integration that looks perfect: batches vanish, the manifold is smooth, and a rare population that was genuinely there has been folded into its nearest common neighbor. You will not find it by admiring the embedding. You find it by checking whether marker genes still separate the populations you expected, whether a known rare type still forms its own cluster, and whether your bio-conservation score dropped when your batch score climbed.
A useful discipline: before integrating, write down the cell types and rare states you expect to recover. After integrating, confirm each one survived. Integration that improves your batch metric while quietly deleting an item from that list has not helped you — it has cost you the finding you ran the experiment for.
Batch correction is one of the few steps in a single-cell pipeline where the prettier result is often the more suspicious one. Treat a suspiciously clean UMAP the way you would treat a variant call with no supporting reads: as a claim to verify, not a conclusion to trust.
Sources
- Luecken MD, et al. Benchmarking atlas-level data integration in single-cell genomics. Nature Methods, 2022. nature.com ↩
- Tran HTN, et al. A benchmark of batch-effect correction methods for single-cell RNA sequencing data. Genome Biology, 2020. ncbi.nlm.nih.gov ↩
- Korsunsky I, et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nature Methods, 2019. nature.com ↩
- Büttner M, et al. A test metric for assessing single-cell RNA-seq batch correction (kBET). Nature Methods, 2019. nature.com ↩