This is something I have been working on this year quite a bit after reading this fabulous resource over the break. It gave me a lot of insight into the various steps of single-cell dataset analysis.
Some highlighted improvements:
faster, less RAM, less storage
rational approach to selecting number of PCs for dimensionality reduction. So far has led to discovery of some clusters that were previously obscured - e.g. dendritic cells in one dataset
interesting discussing on selecting genes for PCA to bias the clustering (e.g. could use genes that distinguish CD8 T-cell subsets for PCA if particularly interested in exploring clusters related to this)
using TSNE instead of UMAP (preserves local relationships at expense of global - opposite of UMAP) which allows spreads local subclusters better
multiBatchNorm to scale for sequencing depth across samples when integrating
more conservative integration by preserving original clusters with clusterMNN seems to handle cases where a cell type doesn't exist in a sample better
guideline for ambient genes is pretty much what I came up with as well (detect and omit). Provides an example of how attempts to adjust for ambient expression can lead to faulty conclusions.
This is something I have been working on this year quite a bit after reading this fabulous resource over the break. It gave me a lot of insight into the various steps of single-cell dataset analysis.
Some highlighted improvements:
multiBatchNormto scale for sequencing depth across samples when integratingclusterMNNseems to handle cases where a cell type doesn't exist in a sample better