Simulating Single-Cell Gene Expression Count Data with Preserved Gene Correlations by scDesign2

Abstract

scDesign2 is a transparent simulator that generates high-fidelity single-cell gene expression count data with gene correlations captured. This article shows how to download and install the scDesign2 R package, how to fit probabilistic models (one per cell type) to real data and simulate synthetic data from the fitted models, and how to use scDesign2 to guide experimental design and benchmark computational methods. Finally, a note is given about cell clustering as a preprocessing step before model fitting and data simulation.

1. BACKGROUND

In the burgeoning field of single-cell transcriptomics, a pressing challenge is to benchmark various experimental protocols and numerous computational methods in an unbiased manner. Although dozens of simulators had been developed for single-cell RNA-seq (scRNA-seq) data, they lacked the capacity to simultaneously achieve the following three goals: preserving genes, capturing gene correlations, and generating any number of cells with varying sequencing depths. To fill in this gap, we developed a new simulator scDesign2 (Sun et al., 2021), which advanced our previous simulator scDesign (Li and Li, 2019), to achieve all three goals. Notably, scDesign2 can generate high-fidelity synthetic data of multiple scRNA-seq protocols and other single-cell gene expression count-based technologies.

This article provides a brief guide to the scDesign2 R package. For help troubleshooting or to provide feedback, please submit an issue to the GitHub page, which contains more documentation.

2. INSTALLATION

The required R version is no earlier than version 3.6.3. To install the scDesign2 package, users can run the following code in R.

if(!require(devtools)) install.packages(“devtools”); library(devtools);

devtools::install_github(“JSB-UCLA/scDesign2”);

To use the package after the installation, users can run

library(scDesign2);

3. MODEL FITTING AND DATA SIMULATION

The input of scDesign2 is a real single-cell gene expression count matrix, where each row represents a gene, each column a cell, and each entry the expression level of a gene in a cell. In addition, each column needs to be labeled with the cell type that the cell belongs to. Based on this count matrix, scDesign2 would first fit one parametric probabilistic model for each cell type and then use the fitted models to simulate data.

In the R package, we have included an example scRNA-seq data set, which profiles the transcriptome of mouse small intestinal epithelial cells (Haber et al., 2017). The file mouse_sie_10x.rds is the full data set, and the file mouse_sie_10x_demo.rds is a data subset containing 1000 genes and 30% cells for demonstration. In the following example code, we will select four cell types from the data subset and perform model fitting and data simulation for each cell type. In scDesign2, the function for model fitting is fit_model_scDesign2(), and the function for data simulation is simulate_count_scDesign2().

• Load data

• data_mat_demo <- readRDS(system.file(“extdata”, “mouse_sie_10x_demo.rds”, package = “scDesign2”));

• Select four cell types; obtain the total cell number and cell type proportions

• cell_type_sel <- c(“Goblet”, “Tuft”, “TA.Early”, “Enterocyte.Progenitor”); data_mat_demo_sel <- data_mat_demo[, colnames(data_mat_demo) %in% cell_type_sel];

• n_cell_old <- ncol(data_mat_demo_sel);

• cell_type_prop <- prop.table(table(colnames(data_mat_demo_sel)));

• Fit models and simulate data for the four cell types (running time within 14 mins on 4 cores)

• RNGkind(“L‘Ecuyer-CMRG”); set.seed(1);

• copula_result <- fit_model_scDesign2(data_mat_demo, cell_type_sel, sim_method = “copula”, ncores = length(cell_type_sel));

• sim_count_copula <- simulate_count_scDesign2(copula_result, sim_method = “copula”,

• n_cell_new = n_cell_old, cell_type_prop = cell_type_prop);

In this example, the selected cell types are in the cell_type_sel vector, the fitted models are in the copula_result object, and the synthetic data set is the sim_count_copula matrix. We set the synthetic data set to have the same total cell number (n_cell_old) and expected cell type proportions (cell_type_prop) as those of the input data matrix data_mat_demo, but users may change the n_cell_new and cell_type_prop arguments in the simulate_count_scDesign2() function.

To evaluate the quality of the synthetic data set, we will combine the synthetic cells with the real cells and examine whether they are indistinguishable in the t-SNE visualization.

if(!require(Rtsne)) install.packages(“Rtsne”); library(Rtsne); set.seed(1); Rtsne_combined <- Rtsne (log(t(cbind(data_mat_demo_sel, sim_count_copula)) +1)); Rtsne_combined_vis <- data.frame(x = Rtsne_combined$Y[, 1], y = Rtsne_combined$Y[, 2], group = factor(c(rep(“real”, ncol(data_mat_demo_sel)), rep(“synthetic”, ncol(sim_count_copula)))), cell_type = factor(c(colnames(data_mat_demo_sel), colnames(sim_count_copula)))); attach(Rtsne_combined_vis); plot(x = x, y = y, pch = c(16, 2)[group], col = c(“red”, “blue”, “green”, “black”)[cell_type]); legend(“topleft”, legend = levels(cell_type), col = c(“red”, “blue”, “green”, “black”), pch = 16, bty = “n”); legend(“bottomright”, legend = c(“real”, “synthetic”), pch = c(21, 2)); detach(Rtsne_combined_vis);

The t-SNE visualization shows that the synthetic cells mix well with the real cells.

4. APPLICATIONS TO EXPERIMENTAL DESIGN AND COMPUTATIONAL BENCHMARKING

Two important applications of scDesign2 are guiding experimental design and benchmarking computational methods. This requires generating synthetic data with varying cell numbers and sequencing depths.

In this study, we demonstrate how to generate synthetic data sets with a fixed total cell number and varying sequencing depths. We will use cell_type_sel, n_cell_old, cell_type_prop, and copula_result from the previous code. The first step is to calculate the sequencing depth of the real data.

total_count_old <- sum(data_mat_demo_sel);

To vary the sequencing depth, we change total_count_old by a factor of 1/8, 1/4, 1/2, 2, 4, or 8. The vector adj_factor contains all the multiplicative factors considered.

adj_factor <- c(1/8, 1/4, 1/2, 1, 2, 4, 8);

Finally, we use the following code for data simulation. In the simulate_count_scDesign2() function, the key arguments include total_count_old, n_cell_old, total_count_new, and n_cell_new. The first two arguments are the sequencing depth and total cell number of the real data, and the last two arguments are the sequencing depth and total cell number of the synthetic data to be generated. To fix the total cell number, we set n_cell_new to n_cell_old; to vary the sequencing depth, we specify total_count_new as total_count_old multiplied by each factor in the adj_factor vector, up to rounding. The list sim_count contains the synthetic data sets, one for each new sequencing depth total_count_new.

set.seed(1); sim_count <- lapply(1:length(adj_factor), function(iter) {simulate_count_scDesign2(copula_result, total_count_old = total_count_old, n_cell_old = n_cell_old, total_count_new = round(adj_factor[iter] * total_count_old), n_cell_new = n_cell_old, cell_type_prop = cell_type_prop, reseq_method = “mean_scale”, cell_sample = TRUE)});

5. A NOTE ON CELL CLUSTERING

The model fitting and data simulation of scDesign2 is performed for each cell type separately. Hence, partitioning cells into cell types is an important preprocessing step of scDesign2. The partitioning can be done based on biological knowledge, for example, cell type marker genes, or by a clustering algorithm, for example, SC3 (Kiselev et al., 2017) or the Louvain algorithm (Blondel et al., 2008).

On the GitHub page, we provide a proof-of-concept demonstration of how to perform cell clustering using the Louvain algorithm in the Seurat package (Stuart et al., 2019) and how to evaluate the clustering result using the ROGUE score (Liu et al., 2020). For scDesign2 users who do not have predefined cell types, they may follow our demonstration to do cell clustering before using scDesign2 to simulate data.

SOFTWARE AVAILABILITY

The scDesign2 R package is released under the MIT License and available at https://github.com/JSB-UCLA/scDesign2.

AUTHOR DISCLOSURE STATEMENT

The authors declare they have no competing financial interests.

FUNDING INFORMATION

This study was supported by the following grants: NSF DBI-1846216 and DMS-2113754, NIGMS R01GM120507 and R35GM140888, Johnson & Johnson WiSTEM2D Award, Sloan Research Fellowship, and UCLA David Geffen School of Medicine W.M. Keck Foundation Junior Faculty Award (to J.J.L.); Rutgers School of Public Health Pilot Grant and NJ ACTS BERD Mini-Methods Grant (to W.V.L.).