Batch effect removal tools

[grst]

Cells should cluster by cell type, not by patient/dataset. We therefore need some way to address batch effects.

Here are some strategies described in the literature:

• Schelker et al. normalize to the average expression of 3559 housekeeping genes to merge three datasets of multiple patients each.
• Papasaikas's and Hemberg's single cell tutorial uses RUVg, a package designed for bulk RNA-seq, to address batch effects in a single dataset. If no spike-ins are available, this boils down to something similar to Schelker's approach. Does not scale for large datasets. Requires to store all data as dense matrix in R (failed for the 250k cells with 192GB of RAM)
• Biscuit applies clustering and batch effect correction in a single step. Used for the Azizi/Peer dataset. Does not support sparse matrices and does not scale for datasets with more than 50k cells.
• Scanorama. Generalization of the MNN approach. Can (without hacks) only account for batch effects between datasets, and not between patients/batches.
• Harmony, novel, scalable batch effect removal tool.
• BBKNN

[FFinotello]

@grst, this section from Hemberg's course might be of interest:

https://hemberg-lab.github.io/scRNA.seq.course/cleaning-the-expression-matrix.html#dealing-with-confounders

[mlist]

@grst https://www.biorxiv.org/content/early/2017/10/27/200345 kBet

[Hoohm]

Yet another one: https://github.com/SydneyBioX/scMerge

[grst]

and yet another one: "RISC" https://www.biorxiv.org/content/biorxiv/early/2018/11/29/483297.full.pdf?%3Fcollection=