Clusters and/or cell types have been identified, we now want to compare sample groups:

• Differential expression - Differences in expression between sample group within a biological state.

• Differential abundance - Differences in cell numbers between sample groups within a biological state.

Replicates are samples not cells:

• single cells within a sample are not independent of each other,
• using cells as replicates amounts to studying variation inside an individual
• while we want to study variation across a population

Pseudo-bulk:

• gene expression levels for each cluster in each sample
• are obtained by summing across cells

Workflow:

• compute pseudo-bulk count by summing across cells,

  • per cluster and per sample

• perform bulk analysis with fewer replicates,

  • for each cluster separately

Method:

• quasi-likelihood (QL) methods from the edgeR package

• negative binomial generalized linear model (NB GLM)

  • to handle overdispersed count data
  • in experiments with limited replication

Steps:

• Remove samples with very low library sizes, e.g. < 20 cells

  • better normalisation

• Remove genes that are lowly expressed,

  • reduces computational work,
  • improves the accuracy of mean-variance trend modelling
  • decreases the severity of the multiple testing correction
  • filter: log-CPM threshold in a minimum number of samples, smallest sample group

• Correct for composition biases

  • by computing normalization factors with the trimmed mean of M-values method

• Test whether the log-fold change between sample groups is significantly different from zero

  • estimate the negative binomial (NB) dispersions
  • estimate the quasi-likelihood dispersions, uncertainty and variability of the per-gene variance

Aim:

• test for significant changes in grouped cell abundance across conditions

Example:

• which cell types are depleted or enriched upon treatment?

• Most methods require defined clusters as input. Assigning cells to discrete clusters in context of continuous differentiation, developmental or stimulation trajectories.

• Methods that don’t require clusters also don’t model variability in cell numbers among replicates or can only carry out pairwise comparisons.

Milo

• Uses K- nearest neighbour graph to model cellular states as overlapping neighbourhoods. Non-independence is accounted for with a weighted version of the Benjamini–Hochberg method.

• Determines neighbourhoods and groupings independently of our defined clusters

• Can be used for complex models

• Faster and scalable

Steps:

• Construct KNN graph

  • rescales UMI count by per-cell sequencing depth
  • log transforms
  • uses PCA
  • calculates Euclidean distance between cells and its k nearest neighbour in PC space

• Defines Cell Neighbourhoods

• Counts cells in Neighbourhoods

• Tests for DA in Neighbourhoods

• Does a multiple testing correction (Spacial FDR)

• Visualiations