Batch correction

[MicTott]

Roadmap for batch correction

Batch correction here is tricky as we need to correct for both Samples within species and then merge the data across Species. For this, I think it makes the most sense to correct for Samples wthin species first, and then merge and correct for differences across the Species datasets.

I will perform testing using both MNN and Seurat as per this paper.

Comparing batch correct methods (Harmony vs MNN vs Seurat)

1 . Correct samples within species

• Remove lowly expressed genes: Keep only those genes that have at least 1 UMI in at least 5% of the data

Within each species...

• Because this is a very large dataset and each dataset have a different number of total cells, let's randomly subset 10k cells from each species. Do not combine subsets.

For Harmony and MNN

• Compute approx. GLM-PCA using null residuals according to this workflow: 'scry::nullResiduals' + 'scater::runPCA with chosen HVGs. n=50 components

• Compute UMAP on using scater::runUMAP on uncorrected GLM-PCA dimred

• Perform batch correction on sample_id using batchelor::reducedMNN with GLM-PCA dimred; add to reducedDims

• Compute UMAP on using scater::runUMAP on MNN corrected GLM-PCA dimred

• Perform batch correction on 'sample_id' using 'Harmony::RunHarmony' with GLM-PCA dimred

• Compute UMAP on using scater::runUMAP on Harmony corrected GLM-PCA dimred

• Visualize batch correction results by plotting and printing uncorrected, MNN, and Harmony corrected UMAPs colored by sample_id

GLM-PCA resulted in "stringy" clusters no matter how I adjusted the hyperparameters. Because I previously determined that Seurat was the best for clustering, I moved forward with that (which uses Pearson residuals from negative binomial regression (sctransform) instead of GLM-PCA)

Seurat CCA

• Perform batch correction using Seurat