phipsonlab / speckle_old

R package for analysing single cell data

speckle

Note: we have created a clean repository of the speckle package in order to submit the package to Bioconductor and will be retiring this version of speckle (v0.99.0). The new repository is https://github.com/phipsonlab/speckle.

The speckle package currently contains functions to analyse differences in cell type proportions in single cell RNA-seq data. As our research into specialised analyses of single cell data continues we anticipate that the package will be updated with new functions.

The propeller, propeller.ttest and propeller.anova functions perform statistical tests for differences in cell type composition in single cell data. In order to test for differences in cell type proportions between multiple experimental conditions at least one of the groups must have some form of biological replication (i.e. at least two samples). For a two group scenario, the absolute minimum sample size is thus three. Since there are many technical aspects which can affect cell type proportion estimates, having biological replication is essential for a meaningful analysis.

The propeller function takes a SingleCellExperiment or Seurat object as input, extracts the relevant cell information, and tests whether the cell type proportions are statistically significantly different between experimental conditions/groups. The user can also explicitly pass the cluster, sample and experimental group information to propeller. P-values and false discovery rates are outputted for each cell type.

The propeller.ttest() and propeller.anova() are more general functions that can account for additional covariates in the analysis.

Installation

If you would like to view the speckle vignette, you can install the released version of speckle from github using the following commands:

# devtools/remotes won't install Suggested packages from Bioconductor
BiocManager::install(c("CellBench", "BiocStyle", "scater"))

remotes::install_github("phipsonlab/speckle", build_vignettes = TRUE, 
dependencies = "Suggest")

In order to view the vignette for speckle use the following command:

browseVignettes("speckle")

If you don't care to view the glorious vignette you can also install speckle as follows:

library(devtools)
devtools::install_github("phipsonlab/speckle")

propeller example

This is a basic example which shows you how to test for differences in cell type proportions in a two group experimental set up.

library(speckle)
library(limma)
library(ggplot2)

# Get some example data which has two groups, three cell types and two 
# biological replicates in each group
cell_info <- speckle_example_data()
head(cell_info)

# Run propeller testing for cell type proportion differences between the two 
# groups
propeller(clusters = cell_info$clusters, sample = cell_info$samples, 
group = cell_info$group)

# Plot cell type proportions
plotCellTypeProps(clusters=cell_info$clusters, sample=cell_info$samples)

Please note that this basic implementation is for when you are only modelling group information. When you have additional covariates that you would like to account for, please use the propeller.ttest() and propeller.anova() functions directly. Please read the vignette for examples on how to model a continuous variable, account for additional covariates and include a random effect in the analysis.

speckle_old's Issues

Fitting a mixed-effect model --> dream() instead of lmFit() or duplicateCorrelation()?

Hi,

Thanks for the clear explanations of your vignette!

I am interested in fitting a mixed-effect model (i.e., including the sequencing lane as a random effect). I noticed that in the 'More complex (or just different) experimental designs' section, you suggest to use lmFit() for LM, or duplicateCorrelation() for LMM, but I was wondering whether I could use dream() on the transformed proportions (props$TransformedProps of the vignette) since dream() can hold mixed-models and lmFit() not. See section 'Dream Analysis' for more details.

Thanks a lot!

Best,
Aida

Error with 2 levels clusters

Hello,

Thank you for this useful package.

When I provide a vector with 2 levels for the clusters argument in the propeller function, I get the following error :
Error in tmixture.matrix(out$t, stdev.unscaled, df.total, proportion, : Dims of tstat and stdev.unscaled don't match

Is it not possible to find differences of proportions when we only have 2 levels of cluster ?

FYI, I don't have any issue when the clusters argument has more than 2 levels and my study is based on the differences between 2 groups.

Thank you for your help.

Regards,
Yannick