Splitting a Count Matrix into Independent Folds.
What is countsplit?
The countsplit R package splits an integer-valued matrix into multiple folds of data with the same dimensions. These folds will be independent under certain modeling assumptions, and can thus be used for cross validation.
For tutorials associated with this package, please visit https://anna-neufeld.github.io/countsplit.tutorials/.
The motivation for this method in the setting where the data are Poisson distributed is described in Neufeld et al., 2022 (link to paper) in the context of inference after latent variable estimation for single cell RNA sequencing data. Briefly, count splitting allows users to perform differential expression analysis to see which genes vary across estimated cell types (such as those obtained via clustering) or along an estimated cellular trajectory (pseudotime). Neufeld et al., 2023 (link to preprint) extends the method to the setting where the data follow a negative binomial distributed, and provides additional settings where count splitting is useful. For example, count splitting is useful broadly for evaluating low-rank representations of the data.
Recent package updates
We recently sped up the performance of the package by re-implementing the main functions in C++. This is especially useful for real scRNA-seq datasets, which are quite large. We would like to acknowledge Mischko Heming (mheming.de) for implementing most of this speedup through a github contribution.
We have consolidated the functions in this package such that both Poisson and negative binomial thinning can be performed using the same function; countsplit. This function can also be used to create an arbitrary number of folds of data, rather than just a single train/test split. If you are a previous user of countsplit, please be sure to read the documentation to see our recent changes!
The vignettes and data associated with this package are stored in the associated ``countsplit.tutorials" package. To see the tutorials, please visit the updated tutorial website: https://anna-neufeld.github.io/countsplit.tutorials/. This change helps with overall package size and build time. Most of the tutorials currently make use of Poisson thinning, but we are in the process of adding more tutorials that use the negative binomial methodology.
How can I get countsplit?
Make sure that remotes is installed by running install.packages("remotes"), then type:
remotes::install_github("anna-neufeld/countsplit")
To also download the data needed to reproduce the package vignettes, be sure to also install the ``countsplit.tutorials" package.
remotes::install_github("anna-neufeld/countsplit.tutorials").
Starting soon, we hope that the countsplit package will be available on CRAN. The countsplit.tutorials package will remain only on github, for size reasons. Future versions of this package will be able to be downloaded with:
install.packages("countsplit")
Where can I learn more?
Please visit our tutorial website https://anna-neufeld.github.io/countsplit.tutorials/ to see an introduction to our framework on simple simulated data, as well as tutorials for integrating the count splitting package with common scRNA-seq analysis pipelines (Seurat, scran, and Monocle3).
Please visit https://github.com/anna-neufeld/countsplit_paper for code to reproduce the figures and tables from our Poisson paper.
Please visit https://github.com/anna-neufeld/nbcs_paper_simulations for code to reproduce the figures and tables from our negative binomial paper.
References
Neufeld, A.,Gao, L., Popp, J., Battle, A. & Witten, D. (2022), ‘Inference after latent variable estimation for single-cell RNA sequencing data’, Biostatistics.
Neufeld, A.,Dharamshi, A., Gao, L., & Witten, D. (2023), ‘Data thinning for convolution-closed distributions’, https://arxiv.org/abs/2301.07276/ .
Neufeld, A., Popp, J., Gao, L., Battle, A. & Witten, D. (2023), ‘Negative binomial count splitting for single-cell RNA sequencing data'. https://arxiv.org/pdf/2307.12985.pdf .