Mapping and batch correction

Compare MNN and Seurat transfer

MNN and seurat transfer

Key differences

• MNN does batch correction by reducedMNNetc. (for each query cell, MNN pair specific batch vector s are weighted by Gaussian kernel, and distance is from the query cell to the cells in the MNN set of query dataset), and then find KNN on corrected space for each query cell to **transfer information.**

  https://static-content.springer.com/esm/art%3A10.1038%2Fnbt.4091/MediaObjects/41587_2018_BFnbt4091_MOESM65_ESM.pdf

  

• When Seurat uses TransferData, it doesn’t do batch effect, but by default uses reference PCA structure to find anchors, and then transfers information using kernel weighted mean.

Workflows

MNN

• Find MNN on raw PCA space
• Correct batch effect, gaussian kernel applied
• Run KNN on corrected space
• Transfer by majority voting among KNN

Seurat

• Find anchors on reference PCA

  Seurat has an option (set by default) to project the PCA structure of a reference onto the query, instead of learning a joint structure with CCA.
  

• Transfer kernel weighted information of anchors

• When IntegrateEmbeddings

  • Correct batch by subtracting kernel weighted anchor-specific correction vectors (the same as MNN)

  https://satijalab.org/seurat/reference/integratedata

  Compute the anchor integration matrix as the difference between the two expression matrices for every pair of anchor cells
  Compute the transformation matrix as the product of the integration matrix and the weights matrix.
  Subtract the transformation matrix from the original expression matrix.
  

  • Project the query data onto the UMAP structure of the reference