PseudotimeDE: inference of differential gene expression along cell pseudotime with well-calibrated p-values from single-cell RNA sequencing data

Abstract

In the investigation of molecular mechanisms underlying cell state changes, a crucial analysis is to identify differentially expressed (DE) genes along a continuous cell trajectory, which can be estimated by pseudotime inference from single-cell RNA-sequencing (scRNA-seq) data. However, existing methods that identify DE genes based on inferred pseudotime do not account for the uncertainty in pseudotime inference. Also, they either have ill-posed p-values that hinder the control of false discovery rate (FDR) or have restrictive models that reduce the power of DE gene identification. To overcome these drawbacks, we propose PseudotimeDE, a robust method that accounts for the uncertainty in pseudotime inference and thus identifies DE genes along cell pseudotime with well-calibrated p-values. PseudotimeDE is flexible in allowing users to specify the pseudotime inference method and to choose the appropriate model for scRNA-seq data. Comprehensive simulations and real-data applications verify that PseudotimeDE provides well-calibrated p-values essential for controlling FDR and downstream analysis and that PseudotimeDE is more powerful than existing methods to identify DE genes.

scDesign2: a transparent simulator that generates high-fidelity single-cell gene expression count data with gene correlations captured

“…In addition to selecting experimental protocols before conducting scRNA-seq experiments, a common challenge after collecting scRNA-seq data is to choose among the many available data analysis methods in an unbiased manner. For example, many algorithms have been developed for missing gene expression imputation [36,37], dimensionality reduction [38][39][40], cell clustering [41][42][43][44], rare cell type detection [45][46][47], differentially expressed gene identification [48][49][50][51][52], and trajectory inference [53][54][55][56][57]. Even though several benchmark and comparative studies have been carried out for common analysis tasks [58][59][60][61][62][63], most of them have only evaluated a subset of available computational methods using data from limited experimental protocols.…”

Clipper: p-value-free FDR control on high-throughput data from two conditions

“…This issue is evidenced by serious concerns about the widespread miscalculation and misuse of p-values in the scientific community [30]. As a result, bioinformatics tools using questionable p-values either cannot reliably control the FDR to a target level [23] or lack power to make discoveries [31]; see Results. Therefore, p-value-free control of FDR is desirable, as it would make data analysis more transparent and thus improve the reproducibility of scientific research.…”

Clipper: p-value-free FDR control on high-throughput data from two conditions

“…However, there is a growing need to generalize our framework to identify features across more than two conditions. For example, temporal analysis of scRNA-seq data aims to identify genes whose expression levels change along cell pseudotime [31]. To tailor Clipper for such analysis, we could define a new contrast score that differentiates the genes with stationary expression (uninteresting features) from the other genes with varying expression (interesting features).…”

“…This issue is evidenced by serious concerns about the widespread miscalculation and misuse of p-values in the scientific community [30]. As a result, bioinformatics tools using questionable p-values either cannot reliably control the FDR to a target level [23] or lack power to make discoveries [31]; see the "Results" section. Therefore, p-value-free control of FDR is desirable, as it would make data analysis more transparent and thus improve the reproducibility of scientific research.…”