The chromVAR R package was originally developed by Alicia Schep and colleagues from the Greenleaf lab (Schep et al. 2017). Although originally designed for single-cell ATAC-seq data, it has been shown to be highly sensitive for bulk data as well (Gerbaldo et al. 2024). The aim of the method is to infer, based on the accessibility of motif matches, the relative activity of transcription factors (TFs) in each sample or cell, adjusting for technical biases (GC content and enrichment bias). It is recommended that you read the chromVAR documentation before using this package.
betterChromVAR is first and foremost a considerably faster, analytical re-implementation of the original method (it is also considerably faster than the C++ reimplementation in ArchR (Granja et al. 2021). Contrarily to the original chromVAR, it is entirely deterministic, and achieves much higher efficiency by replacing permutations with an analytical solution and computing expectations and variance at the level of bias bins, instead of in the peak-space (for more detail, see (Wang et al. 2026)).
In addition, betterChromVAR includes a few additions, such as simpler weighted expectations, bias shrinkage, and an ATAC-seq normalization method based on the chromVAR logic. Importantly, however, not all functionalities of chromVAR have been reimplemented here: betterChromVAR chiefly focuses on the key task of efficiently computing motif deviations.
betterChromVARbetterChromVAR
is a R package available via the Bioconductor repository. It can be
installed using the following commands in your R
session:
betterChromVAR
takes two primary inputs: 1) the counts in peaks across cells or
samples, and 2) an annotation of which TFs/motifs match which peak1. The peak
counts should be provided as a RangedSummarizedExperiment
object2,
while the annotation can either be in that format too or provided as a
(sparse) matrix.
There are multiple ways of generating your peak counts; the original
chromVAR
package includes such a function (getCounts()), and the
epiwraps
package has some with more options (functions
peakCountsFromBAM() and
peakCountsFromFrags()). Similarly, the motifmatchr
package can be used to generate the motif matching annotation. An
important consideration is that the peaks or regions used for the
purpose of this analysis should have similar widths. It is thus highly
recommended, before generating the count and annotation matrices, to
resize your regions (e.g. using
peaks <- resize(peaks, width=300, fix="center")). The
exact size can be something not too large (otherwise the
presence/absence of a motif becomes meaningless) and ideally close to
the median size of your original peaks. In addition, is it advisable to
restrict your peaks to those that are on standard chromosomes3.
Here we’ll use dummy data as example:
suppressPackageStartupMessages({
library(SummarizedExperiment)
library(betterChromVAR)
})
attach(getDummyData())
counts## class: RangedSummarizedExperiment
## dim: 500 10
## metadata(0):
## assays(1): counts
## rownames: NULL
## rowData names(1): bias
## colnames(10): sample1 sample2 ... sample9 sample10
## colData names(0):## 6 x 5 sparse Matrix of class "dgCMatrix"
## motif1 motif2 motif3 motif4 motif5
## [1,] . . . . .
## [2,] . . . . .
## [3,] . . . 1 .
## [4,] . . . 1 1
## [5,] . . . . .
## [6,] . . . . .In this case, we can see that rowData(counts) already
has a bias column indicating the GC content of the
regions:
## DataFrame with 500 rows and 1 column
## bias
## <numeric>
## 1 0.532239
## 2 0.560714
## 3 0.429891
## 4 0.504317
## 5 0.623082
## ... ...
## 496 0.521533
## 497 0.453673
## 498 0.446739
## 499 0.417464
## 500 0.463881Had this not been the case, we would first need to add this using:
where my_genome is a BSgenome
object or similar (e.g. an FaFile4).
Once we have this, we can launch the computation of the deviations :
## class: SummarizedExperiment
## dim: 5 10
## metadata(0):
## assays(2): deviations z
## rownames(5): motif1 motif2 motif3 motif4 motif5
## rowData names(8): N Neff ... var.pval var.adjPval
## colnames(10): sample1 sample2 ... sample9 sample10
## colData names(0):The resulting dev object is a SummarizedExperiment with
the same columns (and colData) as the original counts
object, but with the motifs as rows (instead of the original peaks). The
values can be interpreted as the relative activity, across samples or
cells, of the corresponding TFs5. When the function is used with
fpc=FALSE (and otherwise default arguments) the results are
virtually the same as the original chromVAR (averaging over the noise
coming from the random seeds). fpc=TRUE (default), it
includes an additional finite-population correction which affects motifs
matching a large fraction of the peaks (see the fpc
argument of computeDeviationsAnalytic()).
The object contains two assays: the deviations assay
contains the bias-adjusted deviations from the expectation (by default,
the average), i.e. the difference to the expectation divided by the
expectation, and the z assay contains z-scores, i.e. the
difference to the expectation divided by the variance of the
expectation.
If your samples/cells have similar library sizes, it is recommended
that you use the z assay for downstream analysis (such as
differential TF activity using limma). If the samples have
very different library sizes, the z scores will be
influenced by that, and it might be preferable to use the
deviations assay.
The betterChromVAR() function is actually a wrapper
around three steps, which can also be executed individually for more
customization, or to avoid repeating some computation multiple
times.
The first step is creating the bias bins and their pairwise sampling probabilities. This is achieved with:
## Creating 50*50=2500 bias bins and computing their sampling distances## bcvBackground object with 500 peaks,
## split into 50*50 ( 2500 ) bins.The second step is computing the per-bin background expectations and variances for each sample or cell:
## bcvBackground object with 500 peaks,
## split into 50*50 ( 2500 ) bins.
## Background data filled for 10 samples.The bg object has now been filled with the additional
information. It can now be used for:
## class: SummarizedExperiment
## dim: 5 10
## metadata(0):
## assays(2): deviations z
## rownames(5): motif1 motif2 motif3 motif4 motif5
## rowData names(8): N Neff ... var.pval var.adjPval
## colnames(10): sample1 sample2 ... sample9 sample10
## colData names(0):This last call could be made with different annotations, without needing to recompute the previous steps.
In addition to the enrichment and GC bias taken into account by the
original chromVAR,
betterChromVAR
supports an optional third dimension: fragment length bias (see
?getBackgroundBins() for more information). This requires
the compilation of an additional bias component (the log10-transformed
mean or median fragment length per region). If this is desired, on
option is to use the counting functions from the epiwraps package,
which can provide this information.
If you are using betterChromVAR (or ChromVAR, for that matter) on
single-cell data with multiple cell types, the way to best run it
depends on whether your are interested in differences within cell types,
or between cell types. If interested in differences between cell types,
run it on the entire dataset, specifying the cell types as
grouping argument to betterChromVAR(). In this
way, rare and abundant cell types will be given the same weight in
computing the expectation.
If you are interested in differences within cell types (e.g. between samples/conditions), you should instead run the method separately for each cell type. In this way, the background bins will be defined based on the enrichment in the given cell type, leading to better capture of the bias. The drawback, however, is that the values won’t be comparable across cell types. To test across conditions, we also recommend using it on pseudobulk data.
The only step of the process that needs to be done across the entire
dataset is the computing of the expectation, i.e. the mean (or weighted
mean) for each region across all cells, and the creation of the
background bins, which depends on it. Everything else can easily be
executed in chunks (of cells), as is done for multi-threading in the
betterChromVAR() function.
## Creating 50*50=2500 bias bins and computing their sampling distancesThen we can apply the next steps only on subset of the data:
bg2 <- computeBackgrounds(counts[,1:3], bg, expectation = ex)
dev2 <- computeDeviationsAnalytic(counts[,1:3], bg2, motifMatches)
# this should be identical to what we had run on the whole object:
identical(assay(dev)[,1:3], assay(dev2))## [1] TRUEIf data is on disk, rather than in memory (and you want it to stay
that way), make sure to use this procedure. In recent versions,
getExpectation() should use chunk-based processing. However
it is quite likely that better performance for this step would be
obtained using for instance using the
aggregateAcrossCells() function of the scrapper
package.
The variability of each motif across the dataset is stored in the
rowData of the output of betterChromVAR() :
## DataFrame with 5 rows and 8 columns
## N Neff total variability var.lower var.upper var.pval
## <numeric> <numeric> <numeric> <numeric> <numeric> <numeric> <numeric>
## motif1 81 55.4064 38746 1.11992 0.833634 1.87128 0.140553
## motif2 81 55.7744 44620 1.15434 0.779604 2.12586 0.161801
## motif3 58 39.4625 32727 1.10729 0.870484 1.75189 0.118492
## motif4 70 47.8587 41950 1.06277 0.851162 1.65049 0.157197
## motif5 77 50.5124 34247 1.10142 0.715864 2.10775 0.227652
## var.adjPval
## <numeric>
## motif1 0.202251
## motif2 0.202251
## motif3 0.202251
## motif4 0.202251
## motif5 0.227652This is comparable to the variability produced by the original
chromVAR, however the confidence intervals are analytic
estimates. While these are generally good enough, it is also possible to
obtain bootstrap confidence intervals (as in the original approach)
through the computeMotifVariability() function.
An alternative way to assess how useful or informative motifs are is
how reliably they capture biological variation. This can be estimated by
the proportion of their variance across cells that is not attributable
to sampling noise. The estimateNoiseKappa() function can be
used to assess this:
## Counts assumed to not be paired insertions (`pairs=FALSE`).## kappa reliability reliability.se
## motif1 0.06619570 0.9697782 0.02078492
## motif2 0.11076127 0.9526201 0.04199291
## motif3 0.08796748 0.9593332 0.02528747
## motif4 0.12447380 0.9371495 0.02248975
## motif5 0.08406235 0.9603210 0.05143490While in practice reliability and variability of a motif are correlated, they do not track the same thing, even in this toy example:
## [1] 0.5186066Two motifs can have the same reliability while having very different variability, or the converse. Both are therefore relevant information in selecting useful motifs.
The deviations of individual cells are very noisy: for most motifs,
almost all of the cell-to-cell variation is sampling noise (as estimated
with estimateNoiseKappa()). Working at the pseudobulk level
addresses this issue; alternatively, it is possible to use the
k nearest neighbors (kNN) of a cell to ‘smooth’ its estimates.
smoothDeviationsKNN() does exactly this.
Note that the kNN used for smoothing should not be based on
the motif deviations, but can for example be computed from some
embedding of the peak counts (e.g. LSI or similar).
smoothDeviationsKNN() shrinks the motif deviations of each
cell towards that of its neighbors in an adaptive fashion: cells whose
deviations are more trustworthy (e.g. because the cell has more
fragments) and motifs with higher reliability are smoothed less. The
smoothing assumes that neighboring cells have similar motif activities:
effects that are specific to individual cells and small relative to the
noise are partly smoothed out (the tauScale argument can be
used as a knob to adjust how aggressive the smoothing is).
# for the sake of the example we do a very basic embedding, using a proper
# scATAC-seq embedding is recommended
emb <- prcomp(t(log1p(assay(counts))))$x[,1:3]
knn <- BiocNeighbors::findKNN(emb, k=3)
dev <- smoothDeviationsKNN(dev, knn)
assayNames(dev)## [1] "deviations" "z" "smoothed" "smoothed.z"The smoothed deviations and z-scores are added to the object as the assays. Note that the smoothed values of neighboring cells are not independent observations anymore, and as such they should not be used in tests that treat cells as independent (the kind of tests you should arguably never do anyway).
The general chromVAR approach, and in particular the deterministic
version implemented here, is also amenable to be used to normalize GC-
and enrichment bias out of bulk ATAC-seq data. This is implemented in
the CVnorm() function. In addition, if given a grouping of
the samples (e.g. experimental conditions), the function applies a
variance-based smoothing of the bias correction, inspired by smoothed
quantile normalization (see the qsmooth
package or Hicks et al. (2018) ). In a
nutshell, if the bias in a certain background bin is explained by
experimental groups, it will be less corrected than if it varies across
samples of the groups6.
Example usage:
# we assign arbitrary groups to the samples:
counts$group <- rep(LETTERS[1:2], each=5)
# we run the smoothed CVnorm:
counts <- CVnorm(counts, grouping=counts$group)
# (the normal CVnorm could be run by omitting the grouping)
counts## class: RangedSummarizedExperiment
## dim: 500 10
## metadata(0):
## assays(2): counts corrected
## rownames: NULL
## rowData names(1): bias
## colnames(10): sample1 sample2 ... sample9 sample10
## colData names(1): groupThis adds a corrected assay to the object. Note that
although the assay is corrected for enrichment- and GC-bias (and
fragment length bias if that was provided), it is not corrected for
library size differences. Rather, it is on the original count scale
(although not integer anymore), so that it is amenable to use in
downstream count-based analysis methods such as edgeR.
The scPRINTER python package developed by Ruochi Zhang in the Buenrostro lab includes a GPU-based reimplementation of , with an important modification: rather than cutting the bias space into bins, a continuous, multidimensional bias space space is created from which \(k\) nearest neighbors (kNN) are selected as background. This approach scales nicely to the inclusion of more dimensions of bias. In scPRINTER, the deviations and variance across the \(k\) neighbors is then computed through permutations, as in the classical (although sped up by the GPU).
If a more complex multi-dimensional bias needs to be taken into account, betterChromVAR implements a similar algorithm that combines our analytical framework with the kNN-based approach. An additional feature is that rather than taking the transformed bias space directly to compute distances, we weight the different dimensions by their correlation with peak overdispersion.
The kNN version of betterChromVAR can be used in the following way:
## Computing motif expectations and variances## Calculating final deviations and Z-scoresWe note, however, that this approach is considerably slower (although
still substantially faster than the original chromVAR or the ArchR
re-implementation), requires more memory, and, at the single-cell level,
produces slightly more variable deviations across replicate cells (Wang et al. 2026). We therefore recommend using
the more scalable bin-based variant, unless working with bulk data with
fragment length bias (or any other extra bias dimensions, see the
bias argument of getBackgroundKNN()).
## R version 4.6.1 (2026-06-24)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 26.04.1 LTS
##
## Matrix products: default
## BLAS: /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3
## LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.32.so; LAPACK version 3.12.0
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8
## [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
## [9] LC_ADDRESS=C LC_TELEPHONE=C
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## time zone: Etc/UTC
## tzcode source: system (glibc)
##
## attached base packages:
## [1] stats4 stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] betterChromVAR_1.1.12 SummarizedExperiment_1.43.0
## [3] Biobase_2.73.2 GenomicRanges_1.65.4
## [5] Seqinfo_1.3.2 IRanges_2.47.5
## [7] S4Vectors_0.51.10 BiocGenerics_0.59.12
## [9] generics_0.1.4 MatrixGenerics_1.25.0
## [11] matrixStats_1.5.0 BiocStyle_2.41.0
##
## loaded via a namespace (and not attached):
## [1] Matrix_1.7-6 jsonlite_2.0.0
## [3] crayon_1.5.3 compiler_4.6.1
## [5] BiocManager_1.30.27 Rcpp_1.1.2
## [7] DelayedMatrixStats_1.35.0 Biostrings_2.81.9
## [9] parallel_4.6.1 jquerylib_0.1.4
## [11] BiocParallel_1.47.0 yaml_2.3.12
## [13] fastmap_1.2.0 lattice_0.23-1
## [15] XVector_0.53.0 R6_2.6.1
## [17] S4Arrays_1.13.2 knitr_1.52
## [19] DelayedArray_0.39.8 maketools_1.3.2
## [21] bslib_0.12.0 BiocNeighbors_2.7.3
## [23] rlang_1.3.0 cachem_1.1.0
## [25] xfun_0.62 sass_0.4.10
## [27] sys_3.4.3 otel_0.2.0
## [29] SparseArray_1.13.4 cli_3.6.6
## [31] digest_0.6.39 grid_4.6.1
## [33] sparseMatrixStats_1.25.0 lifecycle_1.0.5
## [35] evaluate_1.0.5 codetools_0.2-20
## [37] buildtools_1.0.0 abind_1.4-8
## [39] rmarkdown_2.32 tools_4.6.1
## [41] htmltools_0.5.9The annotation can be binary or probabilistic (i.e. from 0 to 1). If using motif matches, however, it is recommended to input binary matches, as the magnitude of motif scores is generally poorly correlated to actual binding.↩︎
See SummarizedExperiment if you’re not familiar with those↩︎
See keepStandardChromosomes() from the
GenomeInfoDb
package↩︎
Note, however, that similar motifs will be given similar activity estimates, so that it is often hard to know which of a set of highly-similar motifs is in fact responsible for the observed signal.↩︎
Note that this is different from the original qsmooth
approach – see ?CVnorm for more detail.↩︎