In CompEpigen/scMethrix: Fast and efficient summarization of single cell whole genome bisulfate data

Transformations

Merging experiment objects

During analysis, it might be beneficial to split an experiment into smaller subset experiments, perform operations, and re-assemble the subset experiments back together. Merging can be done by either row or col. This function only affects assays; other experiment components like metadata and reduced dimensionality will be discarded.

m1 <- subset_scMethrix(meth, contigs = "chr1")
m2 <- subset_scMethrix(meth, contigs = "chr2")
meth <- scMethrix::merge_scMethrix(scm1 = m1, scm2 = m2, by="row")

m1 <- subset_scMethrix(meth, samples = "GSM2553093")
m2 <- subset_scMethrix(meth, samples = "GSM2553095")
meth <- scMethrix::merge_scMethrix(scm1 = m1, scm2 = m2, by="col")

Transforming assays

Methylation beta values are presented in different ways based on the upstream pipeline. Scores are commonly reported with ranges such as [-1..1], [0..1], or [0..100]. Depending on your downstream analysis, it may be necessary to transform these values to another range. The transform_assay function allows and arbitrary function input to transform each element in the dataset.

#To convert [0..1] scores to [0..100] scores
fun <- function(x) round(x*100)
meth <- scMethrix::transform_assay(meth, assay = "score" , new_assay = "score.100", trans = fun)

There is an accessible binarize function to transform data into the [0..1] range used by this package for beta values. The function will automatically generate a threshold cutoff based on the mean beta value. This is often useful after imputation.

scm <- scMethrix::transform_assay(scm, assay = "score.100" ,new_assay = "binarize", trans = binarize)

Binning

CpG-level precision may not be necessary for downstream analysis, so binning can be used to drastically reduce the number of calculations necessary for computation. There are multiple options in which to determine binning. The bin_size can be specified for either base pairs (bp) or by the number of CpGs (cpg) per bin. By default, entire chromosomes are used for binning.

#Divide the whole genome into 100kbp bins
scm <- scMethrix::bin_scMethrix(scm, bin_size = 100000, bin_by = "bp")

All assays in the scMethrix object will be binned. By default, the count assay will be summed during binning while the score assay and all other assays will use their mean values. It is possible to specify other operations, if desired.

#Divides genome into groups of 1000 CpGs, but use the median value for the score assay
scm <- scMethrix::bin_scMethrix(scm, bin_size = 1000, bin_by = "cpg", trans = c(score= function(x) median(x,na.rm=TRUE)))

If whole chromosome binning is not desired, a GRanges region list can be provided to bin individual regions of the genome (e.g. promoters), and then re-assemble them into combined object. Typically, one bin will contain one region, but these regions can be further subdivided by bp or cpg sites, as shown before.

# Binning each promoter
if(!requireNamespace("AnnotationHub")) 
  BiocManager::install("AnnotationHub", update = F)

ah = AnnotationHub()
qhs = query(ah, c("RefSeq", "Mus musculus", "mm10"))
genes = qhs[[1]]
proms = promoters(genes)

scm <- scMethrix::bin_scMethrix(scm, regions=proms)

Collapse samples

It may be desirable to collapse multiple samples into a single meta-sample. Using simple grouping, samples can be collapsed via any arbitrary function

# Collapse into simulated cluster_scMethrix() output
colData(scm)["Cluster"] = rep_len(c("Grp1","Grp2"),nrow(colData(scm)))
scm <- scMethrix::collapse_samples(scm, colname, colname = "Cluster")
scm
colData(scm)

# Collapse the samples while using a median opertion for the score assay
colData(scm)["Cluster"] = rep_len(c("Grp1","Grp2"),nrow(colData(scm)))
scm <- scMethrix::collapse_samples(scm, colname, colname = "Cluster", trans = c(score=median))
scm
colData(scm)

Imputation

As single cell bisulfite sequencing is typically very sparse, it is usually necessary to perform some imputation to fill NA elements, as most clustering and dimensionality reduction algorithms cannot handle large numbers of NA values. Imputation algorithms available here are k-nearest neighbour (kNN), iterative PCA (iPCA),or random forest (RF).

#Impute the whole genome via kNN
scm <- scMethrix::impute_regions(scm, assay = "score", new_assay = "impute", type = "kNN", k = 10)
head(score(scm))

By default, the entire chromosome is used for the imputation, but this is incredibly memory-intensive. A GRanges region list can be provided to subset the assay, impute each region individually, then combine them back together. However, there is an assumption for quasi-independence of each region by doing this.

#Impute promoter regions via random forest
scm <- scMethrix::impute_regions(scm, assay = "score", new_assay = "impute", type = "RF", regions = promoters)

For using imputation methods not included in this package, an arbitrary imputation function can be specified. However, the return value must be the imputed matrix with the same dimensions as the input matrix, and in the same column order.

#Impute via an arbitrary imputation function.
fun <- function(mtx) missForest::missForest(mtx, ...)$ximp
scm <- scMethrix::impute_regions(scm, assay = "score", new_assay = "impute", type = fun)

Generating test sets

To judge the accuracy of imputation, it may be helpful to generate a test and training set. This is easily done with generate_training_set. A specified proportion of CpG sites will be subsetted into two scMethrix objects.

#Generate a training set that contains 20% of the total CpG sites.
scm <- scMethrix::generate_training_set(scm, training_prop = 0.2)

For small-scale testing purposes, a randomized miniature scMethrix object may be useful

#Generate a training set that contains 20% of the total CpG sites.
scm <- scMethrix::generate_random_subset(scm, n_cpgs = 10000)

CompEpigen/scMethrix documentation built on Nov. 6, 2021, 3:09 p.m.