• There are generally three reasons for this
  • Batch effects:
    • Process samples in batches, different dates, different technicians, different technologies etc
  • Biological effects:
    • A study involving male and female subjects with the same disease will often have gender-specific clusters when visualized using t-SNE.
    • Need to integrate to remove the “gender” effect and to identify shared cell types.
  • Distinct cellular modalities:
    • For examples for the same study one may profile single cell level transcriptomics or spatial transcriptomics or single cell’s immunophenotype
    • Integration is required to to get comprehensive functional understanding of these data sets.

A few ways our data can be arranged (software-dependent too)

• one large Seurat object containing many samples

• many single-sample Seurat objects, QC’d in isolation

• multiple large Seurat objects with multiple samples

• objects from different software packages (eg. Seurat, SingleCellExperiment, Scanpy)

Important we make sure things match up

• Different bioconductor/package versions

• Different analysts may have formatted things slightly differently

A useful quick look

10X provides software for viewing your cellranger outputs. On a single sample level it can be useful for quick checks but on the output of cellranger aggr it can be valuable to check an experiment has worked before spending large amounts of time with analysis. Especially if you are working in collaboration with a wet lab researcher who may not have the computational skills to do this themselves.

10X also provides a package called LoupeR which can be used as an add on to Seurat to pass filtered or more processed data back for interactive viewing in the Loupe Browser.

• Gaussian/Linear Regression - removeBatchEffect (limma), comBat (sva), rescaleBatches or regressBatches (batchelor)

• Harmony - Korsunsky et al 2019

  • IntegrateLayers (Seurat) or RunHarmony (Harmony)

• Mutual Nearest Neighbours (MNN) correction - Haghverdi et al 2018

  • mnnCorrect (batchelor)

  • FastMNN (batchelor/SeuratWrappers)

• And many more!

  • Different methods may have strengths and weaknesses

  • Benchmark studies can be used as a reference to choose suitable method

PCA embeds cells into a space with reduced dimensionality. Harmony accepts the cell coordinates in this reduced space and runs an iterative algorithm to adjust for dataset specific effects.

• A, Harmony uses fuzzy clustering to assign each cell to multiple clusters, while a penalty term ensures that the diversity of datasets within each cluster is maximized.

• B, Harmony calculates a global centroid for each cluster, as well as dataset-specific centroids for each cluster.

• C, Within each cluster, Harmony calculates a correction factor for each dataset based on the centroids.

• D, Finally, Harmony corrects each cell with a cell-specific factor: a linear combination of dataset correction factors weighted by the cell’s soft cluster assignments made in step a. Harmony repeats steps a to d until convergence. The dependence between cluster assignment and dataset diminishes with each round. Datasets are represented with colors, cell types with different shapes.

Assumptions:

1. The expected input is a matrix normalised for library size.
2. Cells are embedded in low dimensional space (PCA).
3. The low-dimensional nearest-neighbor structure induced by Euclidean distance should be preserved with common similarity metrics such as cosine similarity and correlation. Very unlikely to be violated in real scRNAseq data.

Known Limitations:

1. Performance on small dataset sizes is poor
2. Harmony assumes a linear adjustment is required which may not capture more complex non-linear effects (FastMNN can capture non-linear effects but is more computationally intensive).
3. It’s default correction can be quite harsh so when batch effects are small, parameters need to be changed or a different method used.

• If you use any correction algorithm in the absence of a batch effect, it may not work correctly

• It is possible to remove genuine biological heterogeneity

• In reality the absence of any batch effect would warrant further investigation.

• One way to measure if we have retained heterogeneity is to look at the agreement between clusters before and after correction
• Adjusted Rand Index
• HIGH = GOOD (eg. 0.8 = within batch variation is retained)

The value in batch correction is that it enables you to see population heterogeneity within clusters/celltypes across batches.

• Also increases the number of cells you have

However the corrected values should not be used for gene based analysis eg. DE/marker detection.

• Correction may have introduced artificial agreement between batches on the gene level.

• Integration inherently introduces dependencies between data points which can violate assumptions of statistical tests.