Takes one matrix of counts and splits it into a specified number of folds. Each fold is a matrix of counts with the same dimension as the original matrix. Summing element-wise across the folds yields the original data matrix.

countsplit(X, folds = 2, epsilon = rep(1/folds, folds), overdisps = NULL)

## Arguments

X

A cell-by-gene matrix of integer counts

folds

An integer specifying how many folds you would like to split your data into.

epsilon

A vector, which has length folds, that stores non-zero elements that sum to one. Determines the proportion of information from X that is allocated to each fold. When folds is not equal to 2, the recommended (and default) setting is to allocate equal amounts of information to each fold, such that each element is 1/folds. When folds=2, the default is still (1/2, 1/2), but other values may be beneficial.

overdisps

If NULL, then Poisson count splitting will be performed. Otherwise, this parameter should be a vector of non-negative numbers whose length is equal to the number of columns of X. These numbers are the overdispersion parameters for each column in X. If these are unknown, they can be estimated with a function such as vst in the package sctransform.

## Value

A list of length folds. Each element in the list stores a sparse matrix with the same dimensions as the data X. Each list element is a fold of data.

## Details

When the argument overdisps is set to NULL, this function performs the Poisson count splitting methodology outlined in Neufeld et al. (2022). With this setting, the folds of data are independent only if the original data were drawn from a Poisson distribution.

If the data are thought to be overdispersed relative to the Poisson, then we may instead model them as coming from a negative binomial distribution, If we assume that $$X_{ij} \sim NB(\mu_{ij}, b_j)$$, where this parameterization means that $$E[X_{ij}] = \mu_{ij}$$ and $$Var[X_{ij}] = \mu_{ij} + \mu_{ij}^2/b_j$$, then we should pass in overdisps = $$c(b_1, \ldots, b_j)$$. If this is the correct assumption, then the resulting folds of data will be independent. This is the negative binomial count splitting method of Neufeld et al. (2023).

Please see our tutorials and vignettes for more details.

reference

## Examples

library(countsplit)
library(Matrix)
library(Rcpp)
# A Poisson count splitting example.
n=400
p=2
X <- matrix(rpois(n*p, 7), nrow=n, ncol=p)
split <- countsplit(X, folds=2)
#> As no overdispersion parameters were provided, Poisson count splitting will be performed.
Xtrain <- split[[1]]
Xtest <- split[[2]]
cor(Xtrain[,1], Xtest[,1])
#> [1] 0.06657386
cor(Xtrain[,2], Xtest[,2])
#> [1] -0.02724683

# A negative binomial count splitting example.
X <- matrix(rnbinom(n*p, mu=7, size=7), nrow=n, ncol=p)
split <- countsplit(X, folds=2, overdisps=c(7,7))
Xtrain <- split[[1]]
Xtest <- split[[2]]
cor(Xtrain[,1], Xtest[,1])
#> [1] -0.1417722
cor(Xtrain[,2], Xtest[,2])
#> [1] 0.02038857