Batch correction and data integration in single-cell RNA-seq: Harmony, scVI, Seurat and when to use each
Integrate scRNA-seq samples without erasing biology: choosing Harmony, scVI, scANVI or Seurat, scIB benchmarking and avoiding over-correction.
Key takeaways
- Integration aligns shared cell types across samples. It should not remove real differences between conditions.
- No method wins everywhere: benchmark on your data with batch-removal and biology-conservation metrics.
- Integrated values are for embeddings and clustering. Differential expression uses raw counts with batch in the model.
- A balanced design (each batch contains every condition) matters more than the choice of algorithm.
Any single-cell study with more than one sample has batch effects: differences in capture date, reagent lot, dissociation time, sequencing depth or operator. Integration methods align cells from different samples so that the same cell type clusters together. Done badly, integration either leaves batch structure that looks like biology or removes biology along with the batch.
What integration should and should not do
Should: make T cells from sample 1 and T cells from sample 2 land in the same place, so clustering and annotation reflect cell identity.
Should not: make disease-specific states disappear into control populations, or flatten a real expression shift between conditions.
The trouble is that the algorithm cannot tell batch from biology on its own. Good integration depends on design and on checking.
Start with the design
Batch effects are manageable when every batch contains every condition. They are unfixable when batch and condition are the same thing. If all patient samples were captured on Monday and all controls on Friday, no algorithm can separate "patient" from "Monday" (see Leek and colleagues on the impact of batch effects).
Before choosing a method, we tabulate samples by batch, condition, sex, donor and capture date. If the design is confounded, we say so up front and adjust which claims the data can support.
The main methods
| Method | How it works | Strengths | Watch out for |
|---|---|---|---|
| Harmony | Iteratively corrects PCA embeddings with soft clustering | Very fast, simple, good for moderate effects | Corrects embeddings only, not counts |
| scVI | Variational autoencoder on raw counts, batch as covariate | Handles complex, nested batches; scales to millions of cells | Needs tuning and enough cells per batch |
| scANVI | scVI plus partial cell-type labels | Strong biology conservation; label transfer | Labels must be trustworthy |
| Seurat v5 (CCA/RPCA) | Anchors between datasets | Native in Seurat; RPCA is conservative | CCA can over-correct dissimilar samples |
| BBKNN | Batch-balanced neighbor graph | Fast for huge datasets | Graph only, no corrected embedding |
| Scanorama | Mutual nearest neighbors across datasets | Good on heterogeneous collections | Parameter-sensitive |
In the large benchmark by Luecken and colleagues, deep-learning methods (scANVI, scVI) and Scanorama performed best on complex atlas tasks, while Harmony did well on simpler integration. Tran and colleagues found Harmony, LIGER and Seurat among the strongest on their tasks. The practical lesson is to test rather than assume.
How we benchmark on your data
We run two or three suitable methods and score them with scIB metrics:
- Batch removal: kBET, iLISI, batch ASW, graph connectivity. Do batches mix within cell types?
- Biology conservation: cLISI, cell-type ASW, NMI/ARI against annotations, isolated-label scores. Do cell types stay distinct?
Then we run biological checks that metrics miss. Do disease-only populations stay separate? Do known condition-driven genes still differ? Does any cluster come from just one sample?
What our experts avoid: using integrated or "corrected" expression values for differential expression. Integration outputs are embeddings designed for neighbor graphs. For testing, we go back to raw counts and include batch in the statistical model, usually pseudobulk with a batch covariate.
Special cases
- Multiple chemistries or platforms (3' with 5', cells with nuclei): stronger effects that often need scVI with platform as a covariate, and gene filtering to shared features.
- Mapping to a reference atlas: scArches/scANVI or Azimuth map your cells into an existing atlas without re-integrating it.
- Cross-species: use orthologs and methods designed for it (for example SATURN or SAMap) rather than forcing standard integration.
- Multi-omics: totalVI for RNA plus protein, MultiVI for RNA plus ATAC, and WNN in Seurat. See our multiome guide.
How OmicsQuest can help
Integration benchmarking is part of every multi-sample single-cell analysis we run, and the scIB results go into your QC report so reviewers can see the choice was justified.
Frequently asked questions
Which single-cell integration method is best?
It depends on the data. In the Luecken et al. 2022 benchmark, scANVI, scVI and Scanorama performed well on complex integration tasks, while Harmony performed well on simpler ones and is very fast. We test two or three methods on each dataset and choose using scIB metrics and biological checks.
Can batch correction remove real biology?
Yes. If a condition is confounded with batch, or if integration is too strong, condition-specific cell states can be merged into control populations. That is why we check that known condition-specific populations remain distinct after integration.
Work with OmicsQuest. Related services:
References
- Luecken MD, Buttner M, Chaichoompu K, et al. Benchmarking atlas-level data integration in single-cell genomics. Nat Methods. 2022;19:41-50.
- Tran HTN, Ang KS, Chevrier M, et al. A benchmark of batch-effect correction methods for single-cell RNA sequencing data. Genome Biol. 2020;21:12.
- Korsunsky I, Millard N, Fan J, et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat Methods. 2019;16:1289-1296.
- Lopez R, Regier J, Cole MB, Jordan MI, Yosef N. Deep generative modeling for single-cell transcriptomics. Nat Methods. 2018;15:1053-1058.
- Hao Y, Stuart T, Kowalski MH, et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat Biotechnol. 2024;42:293-304.
- Leek JT, Scharpf RB, Bravo HC, et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat Rev Genet. 2010;11:733-739.
- Heumos L, Schaar AC, Lance C, et al. Best practices for single-cell analysis across modalities. Nat Rev Genet. 2023;24:550-572.