This repository contains all the necessary code to perform the evaluation of differential expression analysis methods in single-cell RNA-seq data, available in
- C Soneson & MD Robinson: Bias, robustness and scalability in single-cell differential expression analysis. Nature Methods 15:255-261 (2018).
In this paper, we compare the performance of more than 30 approaches to differential gene expression analysis in the context of single-cell RNA-seq data. The main results can be further browsed in a shiny app.
Note: The purpose of the conquer_comparison repository is to provide a public record of the exact code that was used for our publication (Soneson & Robinson, Nature Methods 2018). In particular, it is not intended to be a software package or a general pipeline for differential expression analysis of single-cell data. As a consequence, running the code requires the same software and package versions that were used for our analyses (all versions are indicated in the paper). As the analysis involved running a large number of methods on many data sets and over an extended period of time, we cannot guarantee that it will run successfully with new releases of the software, or that exactly the same results will be obtained with newer versions of the packages. While the repository will not be updated to ensure that it runs with every new version of the used packages, the issues can be used to post questions and/or solutions as they arise.
The repository contains the following information:
config/contains configuration files for all the data sets that we considered. The configuration files detail the cell populations that were compared, as well as the number of cells per group used in each comparison.data/contains some of the raw data that was used for the comparison. All data sets that were used can be downloaded as a bundle from http://imlspenticton.uzh.ch/robinson_lab/conquer_de_comparison/export_results/contains results for the final figures, in tabular formatscripts/contains all R scripts used for the evaluationshiny/contains the code for a shiny app built to browse the results (http://imlspenticton.uzh.ch:3838/scrnaseq_de_evaluation)unit_tests/contains unit tests that were used to check the calculationsMakefileis the master script, which outlines the entire evaluation and calls all scripts in the appropriate orderinclude_filterings.mk,include_datasets.mk,include_methods.mkandplot_methods.mkare additional makefiles listing the filter settings, data set and differential expression methods used in the comparison
Assuming that all prerequisites are available, the comparison can be run by simply typing
$ make
from the top directory (note, however, that this will take a significant amount of time!). The Makefile reads the three files include_filterings.mk, include_datasets.mk and include_methods.mk and performs the evaluation using the data sets, methods and filterings defined in these. The file plot_methods.mk detail the methods included in the final summary plots. For the code to execute properly, an .rds file containing a MultiAssayExperiment object for each data set must be provided in the data/ directory. Such files can be downloaded, e.g., from the conquer database. The files used for the evaluation are bundled together in an archive that can be downloaded from here
To add a differential expression method to the evaluation, construct a script in the form of the provided apply_*.R scripts (in the scripts/ directory), where * should be the name of the method. Then add the name of the method to include_methods.mk. To make it show up in the summary plots, add it to plot_methods.mk and assign it a color in scripts/plot_setup.R.
To add a data set, put the .rds file containing the MultiArrayExperiment object in the data/ folder and construct a script in the form of the provided generate_config_*.R scripts (in the scripts/ directory), where * should be the name of the data set. Then add the name of the dataset to the appropriate variables in include_datasets.mk. Also, add the data set to the data/dataset_type.txt file, indicating the type of values in each data set.
Most data sets in the published evaluation are obtained from the conquer repository. The RPM values for the Usoskin dataset was downloaded from http://linnarssonlab.org/drg/ on December 18, 2016. The 10X data set was downloaded from https://support.10xgenomics.com/single-cell-gene-expression/datasets on September 17, 2017.
The list of mouse cell cycle genes was obtained from http://www.sabiosciences.com/rt_pcr_product/HTML/PAMM-020A.html on March 9, 2017.
To run all the unit tests, start R, load the testthat package and run
source("scripts/run_unit_tests.R"). Alternatively, to run just the unit tests in a given file, do e.g. test_file("unit_tests/test_trueperformance.R", reporter = "summary").