This vignette is a condensed version of the documentation pages on the Cicero website. Please check out the website for more details.
The main purpose of Cicero is to use single-cell chromatin accessibility data to predict regions of the genome that are more likely to be in physical proximity in the nucleus. This can be used to identify putative enhancer-promoter pairs, and to get a sense of the overall stucture of the cis-architecture of a genomic region.
Because of the sparsity of single-cell data, cells must be aggregated by similarity to allow robust correction for various technical factors in the data.
Ultimately, Cicero provides a “Cicero co-accessibility” score between -1 and 1 between each pair of accessible peaks within a user defined distance where a higher score indicates higher co-accessibility.
In addition, the Cicero package provides an extension toolkit for analyzing single-cell ATAC-seq experiments using the framework provided by the single-cell RNA-seq analysis software, Monocle. This vignette provides an overview of a single-cell ATAC-Seq analysis workflow with Cicero. For further information and more options, please see the manual pages for the Cicero R package, the Cicero website and our publications.
Cicero can help you perform two main types of analysis:
if (!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager") BiocManager::install("cicero")
Or install the development version of the package from Github.
Cicero holds data in objects of the CellDataSet (CDS) class. The class is derived from the Bioconductor ExpressionSet class, which provides a common interface familiar to those who have analyzed microarray experiments with Bioconductor. Monocle provides detailed documentation about how to generate an input CDS here.
To modify the CDS object to hold chromatin accessibility rather than expression data, Cicero uses peaks as its feature data fData rather than genes or transcripts. Specifically, many Cicero functions require peak information in the form chr1_10390134_10391134. For example, an input fData table might look like this:
The Cicero package includes a small dataset called cicero_data as an example.
For convenience, Cicero includes a function called make_atac_cds. This function takes as input a data.frame or a path to a file in a sparse matrix format. Specifically, this file should be a tab-delimited text file with three columns. The first column is the peak coordinates in the form “chr10_100013372_100013596”, the second column is the cell name, and the third column is an integer that represents the number of reads from that cell overlapping that peak. The file should not have a header line.
The output of make_atac_cds is a valid CDS object ready to be input into downstream Cicero functions.
input_cds <- make_atac_cds(cicero_data, binarize = TRUE)
Because single-cell chromatin accessibility data is extremely sparse, accurate estimation of co-accessibility scores requires us to aggregate similar cells to create more dense count data. Cicero does this using a k-nearest-neighbors approach which creates overlapping sets of cells. Cicero constructs these sets based on a reduced dimension coordinate map of cell similarity, for example, from a tSNE or DDRTree map.
You can use any dimensionality reduction method to base your aggregated CDS on. We will show you how to create two versions, tSNE and DDRTree (below). Both of these dimensionality reduction methods are available from Monocle (and loaded by Cicero).
Once you have your reduced dimension coordinate map, you can use the function make_cicero_cds to create your aggregated CDS object. The input to make_cicero_cds is your input CDS object, and your reduced dimension coordinate map. The reduced dimension map reduced_coordinates should be in the form of a data.frame or a matrix where the row names match the cell IDs from the pData table of your CDS. The columns of reduced_coordinates should be the coordinates of the reduced dimension object, for example:
Here is an example of both dimensionality reduction and creation of a Cicero CDS. Using Monocle as a guide, we first find tSNE coordinates for our input_cds:
set.seed(2017) input_cds <- detectGenes(input_cds) input_cds <- estimateSizeFactors(input_cds) input_cds <- reduceDimension(input_cds, max_components = 2, num_dim=6, reduction_method = 'tSNE', norm_method = "none")
For more information on the above code, see the Monocle website section on clustering cells.
Next, we access the tSNE coordinates from the input CDS object where they are stored by Monocle and run make_cicero_cds:
tsne_coords <- t(reducedDimA(input_cds)) row.names(tsne_coords) <- row.names(pData(input_cds)) cicero_cds <- make_cicero_cds(input_cds, reduced_coordinates = tsne_coords) #> Overlap QC metrics: #> Cells per bin: 50 #> Maximum shared cells bin-bin: 44 #> Mean shared cells bin-bin: 12.3315789473684 #> Median shared cells bin-bin: 1 #> Warning in make_cicero_cds(input_cds, reduced_coordinates = tsne_coords): On #> average, more than 10% of cells are shared between paired bins.
The main function of the Cicero package is to estimate the co-accessiblity of sites in the genome in order to predict cis-regulatory interactions. There are two ways to get this information:
The easiest way to get Cicero co-accessibility scores is to run run_cicero. To run run_cicero, you need a cicero CDS object (created above) and a genome coordinates file, which contains the lengths of each of the chromosomes in your organism. The human hg19 coordinates are included with the package and can be accessed with data(“human.hg19.genome”). Here is an example call, continuing with our example data:
data("human.hg19.genome") sample_genome <- subset(human.hg19.genome, V1 == "chr18") sample_genome$V2 <- 10000000 conns <- run_cicero(cicero_cds, sample_genome, sample_num = 2) # Takes a few minutes to run #>  "Starting Cicero" #>  "Calculating distance_parameter value" #>  "Running models" #>  "Assembling connections" #>  "Successful cicero models: 39" #>  "Other models: " #> #> Zero or one element in range #> 1 #>  "Models with errors: 0" #>  "Done" head(conns) #> Peak1 Peak2 coaccess #> 1 chr18_10006196_10006822 chr18_9755702_9755970 0.00000000 #> 2 chr18_10006196_10006822 chr18_9756925_9757590 -0.12596836 #> 3 chr18_10006196_10006822 chr18_9771216_9771842 -0.17952784 #> 4 chr18_10006196_10006822 chr18_9781976_9782901 -0.06446212 #> 5 chr18_10006196_10006822 chr18_9784605_9785105 0.00000000 #> 6 chr18_10006196_10006822 chr18_9787597_9788029 0.00000000
The Cicero package includes a general plotting function for visualizing co-accessibility called plot_connections. This function uses the Gviz framework for plotting genome browser-style plots. We have adapted a function from the Sushi R package for mapping connections. plot_connections has many options, some detailed in the Advanced Visualization section on the Cicero website, but to get a basic plot from your co-accessibility table is quite simple:
data(gene_annotation_sample) plot_connections(conns, "chr18", 8575097, 8839855, gene_model = gene_annotation_sample, coaccess_cutoff = .25, connection_width = .5, collapseTranscripts = "longest" )