trainSingleR: Train the SingleR classifier

View source: R/trainSingleR.R

trainSingleRR Documentation

Train the SingleR classifier


Train the SingleR classifier on one or more reference datasets with known labels.


  genes = "de",
  sd.thresh = NULL,
  de.method = c("classic", "wilcox", "t"),
  de.n = NULL,
  de.args = list(),
  aggr.ref = FALSE,
  aggr.args = list(),
  recompute = TRUE,
  restrict = NULL,
  assay.type = "logcounts",
  check.missing = TRUE,
  approximate = FALSE,
  num.threads = bpnworkers(BPPARAM),
  BPPARAM = SerialParam()



A numeric matrix of expression values where rows are genes and columns are reference samples (individual cells or bulk samples). Each row should be named with the gene name. In general, the expression values are expected to be log-transformed, see Details.

Alternatively, a SummarizedExperiment object containing such a matrix.

Alternatively, a list or List of SummarizedExperiment objects or numeric matrices containing multiple references, in which case the row names are expected to be the same across all objects.


A character vector or factor of known labels for all samples in ref.

Alternatively, if ref is a list, labels should be a list of the same length. Each element should contain a character vector or factor specifying the label for the corresponding entry of ref.


A string containing "de", indicating that markers should be calculated from ref. For back compatibility, other string values are allowed but will be ignored with a deprecation warning.

Alternatively, if ref is not a list, genes can be either:

  • A list of lists of character vectors containing DE genes between pairs of labels.

  • A list of character vectors containing marker genes for each label.

If ref is a list, genes can be a list of length equal to ref. Each element of the list should be one of the two above choices described for non-list ref, containing markers for labels in the corresponding entry of ref.


Deprecated and ignored.


String specifying how DE genes should be detected between pairs of labels. Defaults to "classic", which sorts genes by the log-fold changes and takes the top de.n. Setting to "wilcox" or "t" will use Wilcoxon ranked sum test or Welch t-test between labels, respectively, and take the top de.n upregulated genes per comparison.


An integer scalar specifying the number of DE genes to use when genes="de". If de.method="classic", defaults to 500 * (2/3) ^ log2(N) where N is the number of unique labels. Otherwise, defaults to 10.


Named list of additional arguments to pass to pairwiseTTests or pairwiseWilcox when de.method="wilcox" or "t".


Logical scalar indicating whether references should be aggregated to pseudo-bulk samples for speed, see aggregateReference.


Further arguments to pass to aggregateReference when aggr.ref=TRUE.


Deprecated and ignored.


A character vector of gene names to use for marker selection. By default, all genes in ref are used.


An integer scalar or string specifying the assay of ref containing the relevant expression matrix, if ref is a SummarizedExperiment object (or is a list that contains one or more such objects).


Logical scalar indicating whether rows should be checked for missing values (and if found, removed).


Logical scalar indicating whether a faster approximate method should be used to compute the quantile.


Integer scalar specifying the number of threads to use for index building.


Deprecated and ignored.


A BiocParallelParam object specifying how parallelization should be performed. Relevant for marker detection if genes = NULL, aggregation if aggr.ref = TRUE, and NA checking if check.missing = TRUE.


This function uses a training data set to select interesting features and construct nearest neighbor indices in rank space. The resulting objects can be re-used across multiple classification steps with different test data sets via classifySingleR. This improves efficiency by avoiding unnecessary repetition of steps during the downstream analysis.

The automatic marker detection (genes="de") identifies genes that are differentially expressed between labels. This is done by identifying the median expression within each label, and computing differences between medians for each pair of labels. For each label, the top de.n genes with the largest differences compared to another label are chosen as markers to distinguish the two labels. The expression values are expected to be log-transformed and normalized.

If restrict is specified, ref is subsetted to only include the rows with names that are in restrict. Marker selection and all subsequent classification will be performed using this restrictive subset of genes. This can be convenient for ensuring that only appropriate genes are used (e.g., not pseudogenes or predicted genes).


For a single reference, a List is returned containing:


An external pointer to various indices in C++ space. Note that this cannot be serialized and should be removed prior to any saveRDS step.


The reference expression matrix. This may have fewer columns than the input ref if aggr.ref = TRUE.


A list containing unique, a character vector of all marker genes used in training; and full, a list of list of character vectors containing the markers for each pairwise comparison between labels.


A list containing unique, a character vector of all unique reference labels; and full, a character vector containing the assigned label for each column in ref.

For multiple references, a List of Lists is returned where each internal List corresponds to a reference in ref and has the same structure as described above.

Custom feature specification

Rather than relying on the in-built feature selection, users can pass in their own features of interest to genes. The function expects a named list of named lists of character vectors, with each vector containing the DE genes between a pair of labels. For example:

genes <- list(
   A = list(A = character(0), B = "GENE_1", C = c("GENE_2", "GENE_3")),
   B = list(A = "GENE_100", B = character(0), C = "GENE_200"),
   C = list(A = c("GENE_4", "GENE_5"), B = "GENE_5", C = character(0))

If we consider the entry genes$A$B, this contains marker genes for label "A" against label "B". That is, these genes are upregulated in "A" compared to "B". The outer list should have one list per label, and each inner list should have one character vector per label. (Obviously, a label cannot have markers against itself, so this is just set to character(0).)

Alternatively, genes can be a named list of character vectors containing per-label markers. For example:

genes <- list(
     A = c("GENE_1", "GENE_2", "GENE_3"),
     B = c("GENE_100", "GENE_200"),
     C = c("GENE_4", "GENE_5")

The entry genes$A represent the genes that are upregulated in A compared to some or all other labels. This allows the function to handle pre-defined marker lists for specific cell populations. However, it obviously captures less information than marker sets for the pairwise comparisons.

If genes is manually passed, ref can be the raw counts or any monotonic transformation thereof. There is no need to supply (log-)normalized expression values for the benefit of the automatic marker detection. Similarly, for manual genes, de.n and sd.thresh have no effect.

Dealing with multiple references

The default SingleR policy for dealing with multiple references is to perform the classification for each reference separately and combine the results (see ?combineRecomputedResults for an explanation). To this end, if ref is a list with multiple references, marker genes are identified separately within each reference if genes = NULL. Rank calculation and index construction is then performed within each reference separately. The result is identical to lapplying over a list of references and runing trainSingleR on each reference.

Alternatively, genes can still be used to explicitly specify marker genes for each label in each of multiple references. This is achieved by passing a list of lists to genes, where each inner list corresponds to a reference in ref and can be of any format described in “Custom feature specification”. Thus, it is possible for genes to be - wait for it - a list (per reference) of lists (per label) of lists (per label) of character vectors.

Note on single-cell references

The default marker selection is based on log-fold changes between the per-label medians and is very much designed with bulk references in mind. It may not be effective for single-cell reference data where it is not uncommon to have more than 50% zero counts for a given gene such that the median is also zero for each group. Users are recommended to either set de.method to another DE ranking method, or detect markers externally and pass a list of markers to genes (see Examples).

In addition, it is generally unnecessary to have single-cell resolution on the reference profiles. We can instead set aggr.ref=TRUE to aggregate per-cell references into a set of pseudo-bulk profiles using aggregateReference. This improves classification speed while using vector quantization to preserve within-label heterogeneity and mitigate the loss of information. Note that any aggregation is done after marker gene detection; this ensures that the relevant tests can appropriately penalize within-label variation. Users should also be sure to set the seed as the aggregation involves randomization.


Aaron Lun, based on the original SingleR code by Dvir Aran.

See Also

classifySingleR, where the output of this function gets used.

combineRecomputedResults, to combine results from multiple references.

rebuildIndex, to rebuild the index after external memory is invalidated.


# Making up some data for a quick demonstration.
ref <- .mockRefData()

# Normalizing and log-transforming for automated marker detection.
ref <- scuttle::logNormCounts(ref)

trained <- trainSingleR(ref, ref$label)

# Alternatively, computing and supplying a set of label-specific markers.
by.t <- scran::pairwiseTTests(assay(ref, 2), ref$label, direction="up")
markers <- scran::getTopMarkers(by.t[[1]], by.t[[2]], n=10)
trained <- trainSingleR(ref, ref$label, genes=markers)

LTLA/SingleR documentation built on July 30, 2022, 4:11 a.m.