ADImpute

ADImpute predicts unmeasured gene expression values from single cell RNA-sequencing data (dropout imputation). This R-package combines multiple dropout imputation methods. ADImpute currently supports, by default, DrImpute and SAVER and two novel imputation methods: Baseline and Network. Baseline imputes dropouts with the average quantified expression level of the respective gene across the dataset. Network uses previously learnt regulatory models of gene expression to infer the expression of dropout genes from the expression of other relevant (predictive) genes in the cell. For more details refer to the preprint.

ADImpute was developed in R 4.0.2, under Linux Mint 20, and tested in Linux, OS X and Windows. For further questions, please contact: ana.carolina.leote@uni-koeln.de

Installation

ADImpute requires R version 4.0. The following R commands should allow you to install ADImpute from github (< 10 seconds):

# install.packages("devtools")
devtools::install_github("anacarolinaleote/ADImpute")
library(ADImpute)

Quick start

Imputation with method(s) of choice

The Impute() function allows for dropout imputation with one or more of the supported methods. In order to specify which imputation method(s) to run, pass them via the do argument:

RPM <- ADImpute::NormalizeRPM(ADImpute::demo_data)
imputed <- Impute(data = RPM, do = c("Network"), cores = 2,
    net.coef = ADImpute::demo_net)

Imputation with ensemble

In addition to running different methods on the data, ADImpute can also determine which of these performs best for each gene and perform an "Ensemble" imputation, which combines the best performing methods for different genes. First, evaluate methods using EvaluateMethods to determine the best performing imputation method for each gene. This step sets a fraction of the quantified entries in the input data to zero, applies different imputation methods to the data and compares the imputation results to the original values. This allows ADImpute to determine which method imputes values with the lowest errors for each gene.

RPM <- ADImpute::NormalizeRPM(ADImpute::demo_data)
methods_pergene <- EvaluateMethods(data = RPM,
    do = c("Baseline", "DrImpute", "Network"),
    cores = 2, net.coef = ADImpute::demo_net)

After determining which method performs best for each gene, the imputation can be re-done on the original data and the results of different methods combined into an ensemble:

imputed <- Impute(do = "Ensemble", method.choice = methods_pergene,
    data = RPM, cores = 2, net.coef = ADImpute::demo_net)

Both the method-specific imputations and the final ensemble results are available for further examination.

Determination of biological zeros

Some zeros in the data correspond to genes expressed in the cell, but not captured upon sequencing - the technical dropouts - while others correspond to genes truly not expressed in the cell - the biological zeros. In order to avoid imputation of biological zeros, ADImpute adapts the well-established approach of scImpute for the computation of the probability of each entry to be a technical dropout. A matrix of such probabilities, of the same size as the original data, can be provided by the user, or computed by ADImpute using scImpute's approach, as below. To activate this option, provide a value for true.zero.thr in the call to Impute(), as exemplified below:

imputed <- Impute(do = "Baseline",
    data = RPM,
    cores = 2,
    true.zero.thr = .3)

Imputation of a SingleCellExperiment

ADImpute can also take a SingleCellExperiment object as input. In this case, EvaluateMethods() will result in new internal row metadata being added to the SingleCellExperiment object, with the best performing methods per gene. Impute() results in new assays being added to the object. If true.zero.thr is specified, only the results after setting biological zeros back to zero will be added to the SingleCellExperiment object.

sce <- NormalizeRPM(sce = ADImpute::demo_sce)
sce <- EvaluateMethods(sce = sce)
sce <- Impute(sce = sce)

Additional imputation methods

ADImpute is built in a modular way, which facilitates the addition of custom functions supporting other imputation methods. Two such methods are scImpute and SCRABBLE, with wrapper functions already contained within ADImpute. To call these methods, please follow these steps:

install scImpute and/or SCRABBLE from their github repositories
clone the ADImpute repository
copy the lines below to the file Wrap.R in the source R/ folder of ADImpute, line #309.
re-load ADImpute using devtools::load_all() on ADImpute's folder

# # call to scImpute
if('scimpute' %in% tolower(do)){
    message('Make sure you have previously installed scImpute via GitHub.\n')
    res <- tryCatch(ImputeScImpute(count_path, labeled = is.null(labels),
            Kcluster = cell.clusters, labels = labels, drop_thre = drop_thre,
            cores = cores, type = type, tr.length = tr.length),
        error = function(e){ stop(paste('Error:', e$message,
            '\nTry sourcing the Impute_extra.R file.'))})
    imputed$scImpute <- log2( (res / scale) + pseudo.count)
}

# call to SCRABBLE
if('scrabble' %in% tolower(do)){
    message('Make sure you have previously installed SCRABBLE via GitHub.\n')
    res <- tryCatch(ImputeSCRABBLE(data, bulk),
                    error = function(e) { stop(paste('Error:', e$message,
                        '\nTry sourcing the Impute_extra.R file.'))})
    imputed$SCRABBLE <- log2( (res / scale) + pseudo.count)
    rm(res);gc()
}

After these steps, scImpute and SCRABBLE can be run with EvaluateMethods or Impute() with do = c("scImpute","SCRABBLE").

beyergroup/ADImpute