Compartmap: Direct inference of higher-order chromatin in single cells from scRNA-seq and scATAC-seq

Compartmap extends methods proposed by Fortin and Hansen 2015, Genome Biology (https://genomebiology.biomedcentral.com/articles/10.1186/s13059-015-0741-y) to perform direct inference of higher-order chromatin in single cells from scRNA-seq and scATAC-seq. Originally, Fortin and Hansen demonstrated that chromatin conformation could be inferred from (sc)ATAC-seq, bisulfite sequencing, DNase-seq and methylation arrays, similar to the results provided by HiC at the group level. Thus, in addition to the base information provided by the aforementioned assays, chromatin state could also be inferred.

Here, we propose a method to infer both group and single-cell level higher-order chromatin states from scRNA-seq and scATAC-seq. To accomplish this, we employ a James-Stein estimator (JSE) towards a global or targeted mean, using either chromsome or genome-wide information from scRNA-seq and scATAC-seq. Additionally, due to the sparsity of single-cell data, we employ a bootstrap procedure to quantify the uncertainty associated with the state and boundaries of inferred compartments. The output from compartmap can then be visualized directly, compared with orthogonal assay types, and/or embedded with something like UMAP or t-SNE. Further, to explore the higher-order interacting domains inferred from compartmap, we use a Random Matrix Theory (RMT) approach to resolve the “plaid-like” patterning, similar to what is observed in Hi-C and scHi-C.

Quick start with example data

Input

The expected input for compartmap is a RangedSummarizedExperiment object. These can be built using the built-in function importBigWig() if starting from BigWigs (recommended for scRNA-seq) or from a feature level object like a SingleCellExperiment with the rowRanges slot populated with the GRanges for each feature (see below in the examples).

# As an example for the quick start, we will load an existing example
# See further down if starting from bigWigs or a feature based object
library(compartmap)

# Load in some example scRNA-seq data from K562
# These data are derived from Johnson and Rhodes et. al 2021 STORM-seq
# They have already been TF-IDF transformed with transformTFIDF()
# See the full workflow below for more details
data("k562_scrna_chr14", package = "compartmap")

Inferring higher-order chromatin domains at the group and single-cell level


#### Group level inference ####
#Process chr14 of the example K562 scRNA-seq data and infer higher-order chromatin at 1Mb resolution
k562_compartments <- scCompartments(k562_scrna_chr14,
                                    chr = "chr14",
                                    res = 1e6,
                                    group = TRUE,
                                    bootstrap = FALSE,
                                    genome = "hg19",
                                    assay = "rna")

#### Single-cell level inference ####
# To infer higher-order domains in single cells and quantifying sign coherence with the bootstrapping procedure, you can run:
# Sub-sample to 10 cells as an example
k562_scrna_chr14.sub <- k562_scrna_chr14[,sample(colnames(k562_scrna_chr14),
                                                 size = 10, replace = FALSE)]
k562_compartments.boot <- scCompartments(k562_scrna_chr14.sub,
                                         chr = "chr14",
                                         res = 1e6,
                                         group = FALSE,
                                         bootstrap = TRUE,
                                         num.bootstraps = 10,
                                         genome = "hg19",
                                         assay = "rna")

# Flip the domain sign if the sign coherence is discordant in 80% of the bootstraps
k562_compartments.boot.fix <- fixCompartments(k562_compartments.boot,
                                              min.conf = 0.8)

# Look at the first cell in the GRangesList object
k562_compartments.boot.fix[[1]]
## GRanges object with 89 ranges and 13 metadata columns:
##                             seqnames              ranges strand |        pc
##                                <Rle>           <IRanges>  <Rle> | <numeric>
##     chr14:19000000-19999999    chr14   19000000-19999999      * | 0.5903261
##     chr14:20000000-20999999    chr14   20000000-20999999      * | 0.3999673
##     chr14:21000000-21999999    chr14   21000000-21999999      * | 0.3047879
##     chr14:22000000-22999999    chr14   22000000-22999999      * | 0.1620188
##     chr14:23000000-23999999    chr14   23000000-23999999      * | 0.0192498
##                         ...      ...                 ...    ... .       ...
##   chr14:103000000-103999999    chr14 103000000-103999999      * |  0.162019
##   chr14:104000000-104999999    chr14 104000000-104999999      * |  0.162019
##   chr14:105000000-105999999    chr14 105000000-105999999      * |  0.162019
##   chr14:106000000-106999999    chr14 106000000-106999999      * |  0.162019
##   chr14:107000000-107349539    chr14 107000000-107349539      * |  0.590326
##                             compartments     score boot.open boot.closed
##                              <character> <numeric> <numeric>   <numeric>
##     chr14:19000000-19999999         open 0.5903261         6           4
##     chr14:20000000-20999999         open 0.3999673         6           4
##     chr14:21000000-21999999         open 0.3047879         6           4
##     chr14:22000000-22999999         open 0.1620188         8           2
##     chr14:23000000-23999999         open 0.0192498         9           1
##                         ...          ...       ...       ...         ...
##   chr14:103000000-103999999         open  0.162019         7           3
##   chr14:104000000-104999999         open  0.162019         6           4
##   chr14:105000000-105999999         open  0.162019         6           4
##   chr14:106000000-106999999         open  0.162019         7           3
##   chr14:107000000-107349539         open  0.590326        10           0
##                              conf.est conf.est.upperCI conf.est.lowerCI
##                             <numeric>        <numeric>        <numeric>
##     chr14:19000000-19999999  0.572247         0.832889         0.311604
##     chr14:20000000-20999999  0.572247         0.832889         0.311604
##     chr14:21000000-21999999  0.572247         0.832889         0.311604
##     chr14:22000000-22999999  0.716740         0.954113         0.479368
##     chr14:23000000-23999999  0.788987         1.000000         0.574032
##                         ...       ...              ...              ...
##   chr14:103000000-103999999  0.644493         0.896662         0.392325
##   chr14:104000000-104999999  0.572247         0.832889         0.311604
##   chr14:105000000-105999999  0.572247         0.832889         0.311604
##   chr14:106000000-106999999  0.644493         0.896662         0.392325
##   chr14:107000000-107349539  0.861234         1.000000         0.679113
##                             flip.compartment flip.score flip.conf.est
##                                    <logical>  <numeric>     <numeric>
##     chr14:19000000-19999999            FALSE  0.5903261      0.572247
##     chr14:20000000-20999999            FALSE  0.3999673      0.572247
##     chr14:21000000-21999999            FALSE  0.3047879      0.572247
##     chr14:22000000-22999999            FALSE  0.1620188      0.716740
##     chr14:23000000-23999999            FALSE  0.0192498      0.788987
##                         ...              ...        ...           ...
##   chr14:103000000-103999999            FALSE   0.162019      0.644493
##   chr14:104000000-104999999            FALSE   0.162019      0.572247
##   chr14:105000000-105999999            FALSE   0.162019      0.572247
##   chr14:106000000-106999999            FALSE   0.162019      0.644493
##   chr14:107000000-107349539            FALSE   0.590326      0.861234
##                             flip.conf.est.upperCI flip.conf.est.lowerCI
##                                         <numeric>             <numeric>
##     chr14:19000000-19999999              0.832889              0.311604
##     chr14:20000000-20999999              0.832889              0.311604
##     chr14:21000000-21999999              0.832889              0.311604
##     chr14:22000000-22999999              0.954113              0.479368
##     chr14:23000000-23999999              1.000000              0.574032
##                         ...                   ...                   ...
##   chr14:103000000-103999999              0.896662              0.392325
##   chr14:104000000-104999999              0.832889              0.311604
##   chr14:105000000-105999999              0.832889              0.311604
##   chr14:106000000-106999999              0.896662              0.392325
##   chr14:107000000-107349539              1.000000              0.679113
##   -------
##   seqinfo: 1 sequence from an unspecified genome; no seqlengths

Visualization of inferred chromatin domains

Once the data have been processed at either the group or single-cell level, one can visualize the results using the plotAB function in compartmap. Notably, we can include the confidence intervals and median, chromosome-wide confidence estimate derived from the bootstrap procedure for sign coherence. At 50%, this suggests that estimates are evenly split between open and closed states. This may be due to data sparsity or heterogeneity in the data. One possible approach to resolve this is to increase the number of bootstraps performed if initially set low (e.g. 10). Alternatively, it may be a region that is worth investigating for your data set.


# Plot the "fixed" results in cell 1 from above with plotAB
# Include the confidence intervals and median confidence estimate
plotAB(k562_compartments.boot.fix[[1]], chr = "chr14",
       what = "flip.score", with.ci = TRUE, median.conf = TRUE)


# It is known that sometimes, the domains may be inverted relative to orthogonal data
# This is also true in Hi-C and scHi-C domain inference
# One can invert all domains by setting reverse = TRUE in the plotAB call

# plotAB(k562_compartments.boot.fix[[1]], chr = "chr14",
#        what = "flip.score", with.ci = TRUE, median.conf = TRUE,
#        reverse = TRUE)

Extraction of domain inflections

It is often of interest to extract the chromatin domain inflection points as they transition from “open” to “closed” states to look for nearby CTCF sites, etc. We can accomplish this task using the getDomainInflections function in compartmap.


# Extract single-cell domain inflections
# Domain inflections can be used to look for nearby CTCF sites, etc.
k562_cell_1_inflections <- getDomainInflections(k562_compartments.boot.fix[[1]],
                                                what = "flip.score",
                                                res = 1e6,
                                                chrs = "chr14",
                                                genome = "hg19")
# Show the inflection points
k562_cell_1_inflections
## GRanges object with 10 ranges and 0 metadata columns:
##        seqnames    ranges strand
##           <Rle> <IRanges>  <Rle>
##    [1]    chr14  19000000      *
##    [2]    chr14  25999999      *
##    [3]    chr14  33000000      *
##    [4]    chr14  37999999      *
##    [5]    chr14  43000000      *
##    [6]    chr14  51999999      *
##    [7]    chr14  58000000      *
##    [8]    chr14  82999999      *
##    [9]    chr14  95000000      *
##   [10]    chr14 107349539      *
##   -------
##   seqinfo: 1 sequence from an unspecified genome; no seqlengths

Importing bigWigs as input to compartmap

The currently recommended input files to compartmap for scRNA-seq are single-cell bigWigs, though does work with a feature/counts based object as demonstrated in the next section. Single-cell bigWigs can be generated through several tools, such as deeptools (https://deeptools.readthedocs.io/en/latest/). To import bigWigs, we can use the importBigWig function in compartmap. This will read in a bigWig file and optionally summarize to an arbitrary bin size. The bin size used in the compartmap manuscript was 1kb and is what we do here as well.


# We can import a list of bigWig files and merge them into a RangedSummarizedExperiment object

# list the example bigWigs
bigwigs <- list.files(system.file("extdata", package = "compartmap"),
                      full.names = TRUE)

# generate the 1kb bins
data("hg19.gr", package = "compartmap")
kb_bins <- tileGenome(seqlengths = seqlengths(hg19.gr)["chr14"],
                      tilewidth = 1000,
                      cut.last.tile.in.chrom = TRUE)

# import
bigwigs_lst <- lapply(bigwigs, function(x) {
  importBigWig(x, bins = kb_bins,
               summarize = TRUE, genome = "hg19")
})

# combine
bigwig_rse <- do.call(cbind, bigwigs_lst)
colnames(bigwig_rse) <- c("cell_1", "cell_2")

Starting with a feature or counts-based object

In the cases where we do not have or can’t start with bigWigs (e.g. scATAC), we can start with a feature-level or counts object (e.g. SingleCellExperiment). The two things that must be there are making sure rowRanges and colnames are set for each feature and cell/sample. We will use the scATAC-seq from K562 as an example of how the object should look, but will also show one way to add rowRanges to a SingleCellExperiment, which works the same way for a SummarizedExperiment object.


# load the scATAC-seq example
# these data were pre-processed using the csaw package
data("k562_scatac_chr14", package = "compartmap")

# show the overall object structure
# NOTE that the colnames are also there and the assay name is 'counts'
k562_scatac_chr14
## class: RangedSummarizedExperiment 
## dim: 11404 279 
## metadata(6): spacing width ... param final.ext
## assays(1): counts
## rownames: NULL
## rowData names(0):
## colnames(279): cell_1 cell_2 ... cell_287 cell_288
## colData names(4): bam.files totals ext rlen

Showing the rowRanges slot is a GRanges for each feature

rowRanges(k562_scatac_chr14)
## GRanges object with 11404 ranges and 0 metadata columns:
##           seqnames              ranges strand
##              <Rle>           <IRanges>  <Rle>
##       [1]    chr14   19614351-19614500      *
##       [2]    chr14   19614401-19614550      *
##       [3]    chr14   19614451-19614600      *
##       [4]    chr14   19614501-19614650      *
##       [5]    chr14   19614551-19614700      *
##       ...      ...                 ...    ...
##   [11400]    chr14 106938701-106938850      *
##   [11401]    chr14 106938751-106938900      *
##   [11402]    chr14 106938801-106938950      *
##   [11403]    chr14 107280901-107281050      *
##   [11404]    chr14 107280951-107281100      *
##   -------
##   seqinfo: 1 sequence from an unspecified genome

But if we don’t have rowRanges for something like a SingleCellExperiment we are working with, we need to generate them. Thus, we will show an example of how to add rowRanges from a GTF file to a SingleCellExperiment.


# define a helper function for adding rowRanges
# modified from https://github.com/trichelab/velocessor/blob/master/R/import_plate_txis.R
# NOTE that you can modify the rtracklayer::import to not subset to gene level
# if using feature level information (e.g. a bed file of fragments for scATAC)
getRowRanges <- function(gtf, asys) { 
  gxs <- subset(rtracklayer::import(gtf), type=="gene")
  names(gxs) <- gxs$gene_id
  granges(gxs)[rownames(asys)]
}

# import the example HEK293T SingleCellExperiment from the SMART-seq3 paper
data("ss3_umi_sce", package = "compartmap")

# import the example GTF with the helper function
gtf_rowranges <- getRowRanges(system.file("extdata/grch38_91_hsapiens.gtf.gz",
                                          package = "compartmap"),
                              ss3_umi_sce)

# add rowRanges to the SingleCellExperiment
rowRanges(ss3_umi_sce) <- gtf_rowranges

# we can now proceed to the next steps and run the compartmap workflow

Running the compartmap workflow

Once we have data in a RangedSummarizedExperiment or other type of SummarizedExperiment with the rowRanges slot filled, we can proceed through the compartmap workflow. We will use the same K562 scRNA-seq data shown in the manuscript on chromosome 14 here as the example.


# Load example K562 data imported using importBigWig
data("k562_scrna_raw", package = "compartmap")

# TF-IDF transform the signals
k562_scrna_chr14_tfidf <- transformTFIDF(assay(k562_scrna_se_chr14))

# Add back the TF-IDF counts to the object in the counts slot
assay(k562_scrna_se_chr14, "counts") <- t(k562_scrna_chr14_tfidf)

# Compute chromatin domains at the group level
k562_scrna_chr14_raw_domains <- scCompartments(k562_scrna_se_chr14,
                                               chr = "chr14",
                                               res = 1e6,
                                               group = TRUE,
                                               bootstrap = TRUE,
                                               num.bootstraps = 10,
                                               genome = "hg19",
                                               assay = "rna")

# For single-cells, run the following
# k562_scrna_chr14_raw_domains <- scCompartments(k562_scrna_se_chr14,
#                                                chr = "chr14",
#                                                res = 1e6,
#                                                group = FALSE,
#                                                bootstrap = TRUE,
#                                                num.bootstraps = 10,
#                                                genome = "hg19",
#                                                assay = "rna")

# 'Fix' compartments with discordant sign coherence
k562_scrna_chr14_raw_domains.fix <- fixCompartments(k562_scrna_chr14_raw_domains)

# Plot results
plotAB(k562_scrna_chr14_raw_domains.fix,
       chr = "chr14",
       what = "flip.score",
       with.ci = TRUE,
       median.conf = TRUE)

Higher-order chromatin interaction maps

Another interesting aspect we can derive from scRNA and scATAC is the higher-order interacting domains through denoising of the correlation matrices using a Random Matrix Theory approach. This is often represented with the “plaid-like” patterning shown in Hi-C and scHi-C approaches where stronger correlations (e.g. greater intensity of red) indicates interacting domains relative to lesser correlation. We can do something similar here.


# Iterative denoising using Random Matrix Theory approach
k562_scrna_chr14_rmt <- getDenoisedCorMatrix(k562_scrna_chr14,
                                             res = 1e6,
                                             chr = "chr14",
                                             genome = "hg19",
                                             assay = "rna",
                                             iter = 2)
## Shrinking bins with the JSE.
## Calculating correlations...
## Done...
## Denoising the correlation matrix using RMT.
## Iterative denoising. Iteration: 2

# Plot the interaction map
plotCorMatrix(k562_scrna_chr14_rmt,
              uppertri = TRUE)