SCEnt
SCEnt is a package for single cell entropy analysis. It can calculate metrics for the heterogeneity or homogeneity of a gene within a cell population. It can use these metrics to perform feature selection for scRNA-seq data.
From the work of Michael J. Casey, Jörg Fliege, Ruben J. Sanchez-Garcia and Ben D. MacArthur. Package written by Hugh Warden.
Installation
And the development version of SCEnt is available from GitHub with:
# install.packages("devtools")
devtools::install_github("hwarden162/SCEnt")
Entropy Analysis
The expression of each gene can be considered as a probability distribution. Where given the information that a gene has been expressed, the expression counts can be used to infer the probability that it was expressed in a given cell.
Each of these probability distributions for each gene contiain a certain amount of information, known in mathematics as entropy. Entropy has been historically described as the amount of ‘surprise’ encoded in the system. If a given gene is only ever expressed in one cell then this gene is said to have low entropy, as if that gene is expressed then the cell expressing it is known and there is no ‘surprise’. However, if a given gene is expressed equally in every cell, then this gene will have high entropy. As expression of this gene does not narrow down the amount of cells it could have come from, meaning the answer will be a ‘surprise’.
What this means, is that the more homogeneously a gene is expressed within a cell population, the more entropy the expression distrbution will have. Therefore, allowing the entropy to be a metric of homogeneity within the cell population.
Every probability distribution has an entropy value and there are many ways to compare entropy values between probability distributions. One of these is called the Kullback-Liebler Divergence (or KL Divergence for short). KL Divergence measures the amount of entropy that would be lost if one distribution were used to approximate another.
The KL Divergence can then be used in this case to see how much information is lost if a uniform distribution were used to represent the expression distribution of a gene. Remembering that the uniform distribution would be the expression of a completely homogeneous gene, the KL Divergence gives us a measure of how much extra information we are getting from the heterogeneity of the gene. Essentially giving a metric for the heterogeneity of a gene within a cell population.
This is intended to be a very brief overview of the underpinning mathematics of SCEnt. For a more detailed explanation please see the paper Measuring the Information Obtained from a Single-Cell Sequencing Experiment.
Using SCEnt to Quantify Homogeneity and Heterogeneity
Here is some synthetic scRNA-seq data:
gene_counts
#> gene1 gene2 gene3 gene4 gene5
#> cell1 0 5 2 3 0
#> cell2 0 5 0 3 0
#> cell3 0 3 2 3 0
#> cell4 0 2 1 3 0
#> cell5 1 0 3 3 5
#> cell6 2 0 0 3 0
#> cell7 3 0 1 3 0
The gene_hom() and gene_het() functions can be used to calculate the homogeneity or heterogeneity of a gene, respectively. Each of these can be passed a gene and it will return a value.
(gene1 <- gene_counts[,1])
#> cell1 cell2 cell3 cell4 cell5 cell6 cell7
#> 0 0 0 0 1 2 3
gene_hom(gene1)
#> [1] 1.459148
gene_het(gene1)
#> [1] 1.348207
(gene2 <- gene_counts[,2])
#> cell1 cell2 cell3 cell4 cell5 cell6 cell7
#> 5 5 3 2 0 0 0
gene_hom(gene2)
#> [1] 1.908613
gene_het(gene2)
#> [1] 0.8987422
(gene3 <- gene_counts[,3])
#> cell1 cell2 cell3 cell4 cell5 cell6 cell7
#> 2 0 2 1 3 0 1
gene_hom(gene3)
#> [1] 2.19716
gene_het(gene3)
#> [1] 0.6101952
(gene4 <- gene_counts[,4])
#> cell1 cell2 cell3 cell4 cell5 cell6 cell7
#> 3 3 3 3 3 3 3
gene_hom(gene4)
#> [1] 2.807355
gene_het(gene4)
#> [1] 0
(gene5 <- gene_counts[,5])
#> cell1 cell2 cell3 cell4 cell5 cell6 cell7
#> 0 0 0 0 5 0 0
gene_hom(gene5)
#> [1] 0
gene_het(gene5)
#> [1] 2.807355
Note that gene4 is a maximally homogeneous gene and gene5 is a maximally heterogeneous gene. Therefore, their entropy calculations always return the extreme values of 0 or 2.807355, the non-zero extreme value will be equal to log2[N] where N is the number of cells in the sample.
Rather than submitting each gene individually, the whole matrix of gene expressions can be passed. However, the data needs to be in the format of having genes represented as rows and cells represented as columns.
gene_hom(t(gene_counts))
#> gene1 gene2 gene3 gene4 gene5
#> 1.459148 1.908613 2.197160 2.807355 0.000000
Rather than transposing the matrix as an input, there is also a transpose parameter for convenience
gene_het(gene_counts, transpose = TRUE)
#> gene1 gene2 gene3 gene4 gene5
#> 1.3482070 0.8987422 0.6101952 0.0000000 2.8073549
The gene_hom() and gene_het() functions have other parameters not covered in detail here. The unit parameter is passed to the entropy calculations to change the units by which entropy is calclated, by default this is set to "log2" such that all outputs are in bits. The normalise parameter is there for if the gene expression is already in probability form. This parameter should generally not be used except in some cases to speed up calculations.
Using SCEnt For Feature Selection
When wanting to use gene expression for prediction, homogeneously expressed genes within the population are not going to be useful for any sort of classification. Conversely, heterogeneously expressed genes are much more likely to be useful to make predictions from.
The function scent_select() will carry out feature selection on scRNA-seq data by finding gene heterogeneity and applying some user defined threshold. The thresholds that can be selected are bit_threshold, count_threshold and perc_threshold. bit_threshold takes a numeric value and SCEnt will only select genes with a heterogenity value greater than the given bit value. count_threshold takes an integer value, SCEnt will return the data with only the top n heteogeneously expressed genes, where n is the count_threshold. perc_threshold takes a value between 0 and 1, SCEnt will only return genes with heterogeneity greater than the percentile of population heterogeneity as given by perc_threshold. Unlike the entropy calculations above, the matrix of gene expressions should be in the format of cells as rows and genes as columns. Again, there is a transpose option built in, as well as the unit and normalise parameters to be passed to the entropy calculations.
Here is the sample data again:
gene_counts
#> gene1 gene2 gene3 gene4 gene5
#> cell1 0 5 2 3 0
#> cell2 0 5 0 3 0
#> cell3 0 3 2 3 0
#> cell4 0 2 1 3 0
#> cell5 1 0 3 3 5
#> cell6 2 0 0 3 0
#> cell7 3 0 1 3 0
scent_select() can carry out each of these three forms of feature selection depending on which threshold is supplied.
scent_select(gene_counts, bit_threshold = 0.85)
#> gene1 gene2 gene5
#> cell1 0 5 0
#> cell2 0 5 0
#> cell3 0 3 0
#> cell4 0 2 0
#> cell5 1 0 5
#> cell6 2 0 0
#> cell7 3 0 0
scent_select(gene_counts, count_threshold = 2)
#> gene1 gene5
#> cell1 0 0
#> cell2 0 0
#> cell3 0 0
#> cell4 0 0
#> cell5 1 5
#> cell6 2 0
#> cell7 3 0
scent_select(gene_counts, perc_threshold = 0.25)
#> gene1 gene2 gene3 gene5
#> cell1 0 5 2 0
#> cell2 0 5 0 0
#> cell3 0 3 2 0
#> cell4 0 2 1 0
#> cell5 1 0 3 5
#> cell6 2 0 0 0
#> cell7 3 0 1 0
Trying to threshold on multiple values will throw an error
scent_select(gene_counts, bit_threshold = 0.85, count_threshold = 2)
#> Error in scent_select(gene_counts, bit_threshold = 0.85, count_threshold = 2):
#> Only one threshold can be set at a time
This is to avoid problems with orders of operation, and instead scent_select() should be piped back into itself so that the order in which the thresholds are applied is explicit
gene_counts %>%
scent_select(bit_threshold = 0.85) %>%
scent_select(count_threshold = 2)
#> gene1 gene5
#> cell1 0 0
#> cell2 0 0
#> cell3 0 0
#> cell4 0 0
#> cell5 1 5
#> cell6 2 0
#> cell7 3 0
Tidy Implementation of Feature Selection
There is a tidy wrapper for scent_select() called scent_select_tidy(). This is currently being used to develop a new recipe step for use within the tidymodels framework.