Goals:

• identify genes that differentially expressed between clusters,

• exclusively or not,

• using different methods that test:

  • the mean expression level,

  • the whole distribution,

  • or the proportion of cells expressing the gene

• compile a summary table.

Challenges:

• Over-interpretation of the results

• Combining different types of marker identification

Recommendations:

• Think of the results as hypotheses that need verification.

• Identify all markers conserved between conditions for each cluster

• Identify markers that are differentially expressed between specific clusters

• Differential expression is comparative. Common comparisons include:

• pairwise cluster comparisons,

  • eg. cluster 1 vs cluster2, cluster 2 vs cluster 3 etc.

• for a given cluster finding ‘marker genes’ that are:

  • DE compared to all cells outside of the cluster
  • DE compared to at least one other cluster
  • DE compared to each of the other clusters
  • DE compare to “most” of the other clusters
  • DE and up-regulated (up- regulated markers are somehow easier to interpret)

• cell-type comparisons (if cell type is known) (with and without clustering)

high noise levels (technical and biological factors)

low library sizes

low amoung of available mRNAs result in amplification biases and dropout events

3’ bias, partial coverage, uneven depth

stochastic nature of transcription

multimodality in gene expression (presence of multiple possible cell states within a cell population)

• The better model fits to the data, the better (more accurate) statistics

• When we cannot fit a model to our data, we resort to non-parametric models (e.g. Wilcoxon rank-sum test, Kruskal-Wallis, Kolmogorov-Smirnov test)

• Non-parametric tests generally convert observed expression values to ranks

• They test whether the distribution of ranks for one group are significantly different from the distribution of ranks for the other group

• May fail in presence of large number of tied values, such as the case of dropouts (zeros) in scRNA-seq

• If the conditions for a parametric test hold, then it will be typically more powerful that a non-parametric test

• Wilcoxon rank-sum test: “It is equally likely that a randomly selected cell from cluster 1 will have higher or lower expression of the gene than a randomly selected cell from cluster 2”

• Binomial test: “Probability of being expressed is the same in cluster 1 and cluster 2”

• t-test: “Mean expression of genes in cluster 1 and cluster 2 are the same”

Some highlights:

• t-test and Wilcoxon work well, given at least few dozens cells to compare

• Bulk RNA-seq analysis methods do not generally perform worse than those specifically developed for scRNA-seq

• Filtering out lowly expressed genes in quite important for good performance of bulk methods (edgeR, DEseq2)

• It’s important to understand what are we trying to compare, e.g. mean expressions, or probability of being expressed

• It’s important to understand the data

• It’s important to assess and validate the results

• Always go back to RNA assay (or similar) for doing differential expression.

• Depending on the method you chose use: counts, normalised counts or lognormalized counts.

• Normalization strategy has a big influence on the results in differential expression, size factors may help.

  • e.g comparing celltype with few expressed genes vs a cell type with many genes.

• Do not forget to account for batch effect

  • You can use the block argument in findMarkers to model batch effect

• A lot of what you get will be noise. Take two random set of cells and run DE and you probably with have a few significant genes with most of the commonly used tests.

• “Always go back to RNA assay (or similar) for doing differential expression.”

  • We will use normalized expression values (not corrected) for testing for differential expression between clusters

• “Do not forget to account for batch effect”

  • We will model the batch effect (SampleGroup) using the block command