Background The development of single-cell RNA sequencing (scRNA-seq) has enabled scientists to catalog and probe the transcriptional heterogeneity of individual cells in unprecedented detail. A common step in the analysis of scRNA-seq data is the selection of so-called marker genes, most commonly to enable annotation of the biological cell types present in the sample. In this paper, we benchmark 59 computational methods for selecting marker genes in scRNA-seq data. Results We compare the performance of the methods using 14 real scRNA-seq datasets and over 170 additional simulated datasets. Methods are compared on their ability to recover simulated and expert-annotated marker genes, the predictive performance and characteristics of the gene sets they select, their memory usage and speed, and their implementation quality. In addition, various case studies are used to scrutinize the most commonly used methods, highlighting issues and inconsistencies. Conclusions Overall, we present a comprehensive evaluation of methods for selecting marker genes in scRNA-seq data. Our results highlight the efficacy of simple methods, especially the Wilcoxon rank-sum test, Student’s t -test, and logistic regression.

A common problem in many disciplines is the need to assign a set of items into categories or classes with known labels. This is often done by one or more expert raters, or sometimes by an automated process. If these assignments or `ratings' are difficult to make accurately, a common tactic is to repeat them by different raters, or even by the same rater multiple times on different occasions. We present an R package `rater`, available on CRAN, that implements Bayesian versions of several statistical models for analysis of repeated categorical rating data. Inference is possible for the true underlying (latent) class of each item, as well as the accuracy of each rater. The models are extensions of, and include, the Dawid-Skene model, and we implemented them using the Stan probabilistic programming language. We illustrate the use of `rater` through a few examples. We also discuss in detail the techniques of marginalisation and conditioning, which are necessary for these models but also apply more generally to other models implemented in Stan.

Bayesian analysis methods often use some form of iterative simulation such as Monte Carlo computation. Models that involve discrete variables can sometime pose a challenge, either because the methods used do not support such variables (e.g. Hamiltonian Monte Carlo) or because the presence of such variables can slow down the computation. A common workaround is to marginalise the discrete variables out of the model. While it is reasonable to expect that such marginalisation would also lead to more time-efficient computations, to our knowledge this has not been demonstrated beyond a few specialised models. We explored the impact of marginalisation on the computational efficiency for a few simple statistical models. Specifically, we considered two- and three-component Gaussian mixture models, and also the Dawid-Skene model for categorical ratings. We explored each with two software implementations of Markov chain Monte Carlo techniques: JAGS and Stan. We directly compared marginalised and non-marginalised versions of the same model using the samplers on the same software. Our results show that marginalisation on its own does not necessarily boost performance. Nevertheless, the best performance was usually achieved with Stan, which requires marginalisation. We conclude that there is no simple answer to whether or not marginalisation is helpful. It is not necessarily the case that, when turned 'on', this technique can be assured to provide computational benefit independent of other factors, nor is it likely to be the model component that has the largest impact on computational efficiency.