ebi-gene-expression-group/selectbcm

An R tool to detect and correct batch-effects in gene-expression data (Microarray and bulk-RNAseq))

SelectBCM tool: Short introduction

"Madhulika Mishra, Lucas Barck, Pablo Moreno, Guillaume Heger, Yuyao Song, Janet M. Thornton, Irene Papatheodorou"

Introduction

Integrating information from bulk transcriptomes from different experiments can potentially increase the impact of individual studies in areas such as biomedicine. Despite that essential data sources for basic and disease biology are available, adequate integration remains difficult. Batch-effect correction for both experimental and technological variations is one of the main challenges in transcriptomics studies. To handle this, various batch-correction methods(BCMs) have been developed in recent times. However, the lack of a user-friendly workflow to select the most appropriate BCM for the given set of experiments needs to be resolved for applying batch correction at a large scale e.g. for multi-cohort transcriptomic analysis. In this project, we developed an R tool named “SelectBCM Tool” to increase the reusability and reproducibility of these BCMs, in order to facilitate the application, comparison and selection of BCMs. Our tools aims to aid choosing the most suitable batch correction method for a set of bulk transcriptomic experiments.

Project overview

SelectBCM performes the following analysis:

• Fetching experiments from ExpressionAtlas based on a biological factor of interest, such as disease, species and tissue
• Removing isolated experiments
• Merging experiments in a single dataset
• detetction of batch effect in the merged dataset
• Correcting batch effects
• Evaluation of batch correction methods

The above steps have been implemented in a variety of forms including stand-alone tools, scripts or R package functions. By creating SelectBCM we aim to provide the infrastructure to compare equivalent logical steps between workflows, and to ‘mix and match’ those components for a custom optimal workflow. The current R package provides a flexibility to perform batch correction and evaluation of various batch-correction methods for given dataset. Finally, it provide performance rank of each BCMs for the given dataset.

Installation

It is advised to set up a dedicated conda environment for a new SelectBCM project. To use SelectBCM with conda, run the following commands from a bash shell:

conda env create -n selectBCM -c r r-base=4.2.2
conda activate selectBCM
conda install -c conda-forge r-remotes
R

In the R session, run the following installation commands:

remotes::install_github("tengfei-emory/scBatch", ref="master")
remotes::install_github("ebi-gene-expression-group/selectBCM", ref="master", build=FALSE)

Alternatively, we also created Docker images to run SelectBCM R package inside a Docker container on either linux/amd64 or linux/arm64 platforms. We also provide the Docker file of the SelectBCM Docker images for custom container building from scratch.

linux/amd64:

Pull container from Docker and open an interactive R session:

docker pull yysong123/selectbcm:amd64
docker run -it yysong123/selectbcm:amd64

linux/arm64:

docker pull yysong123/selectbcm:arm64
docker run -it yysong123/selectbcm:arm64

For a renv approach of version control, we provide renv.lock for the package dependencies. Install the specific package versions as recorded in renv.lock using renv::restore() at the local package directory.

SelectBCM can be installed via github-

install_github('https://github.com/ebi-gene-expression-group/selectBCM')

Or download the package as zip archive and install-

install.packages('selectBCM-master.zip', repos = NULL, type = 'source')

Overview of steps available in SelectBCM

SelectBCM package has several steps ranging from meta-experiment creation to batch-effect evaluation step. In the current wrapper, scripts are written in a user-friendly way. Short description of each step and example is given below -

loading library

Sometime loading of 'magrittr','purrr' and 'dplyr' with SelectBCM package is deprecated, therefore, it is recommended to load all of these together.

Recommendation: Computation of scBatch requires high memory allocation for Rcpp, therefore it is advised to increase R memory.

library(DESeq2)
library(SelectBCM)
library(magrittr)
library(dplyr)
library(purrr)
library(tibble)
library(gtools)
library(readr)
library(Biobase)

library(ggplot2)
library(scales)
library(ggpubr)

Step1: Getting data

Loading local data in R: If the data files are already on the user's local machine, they can be used directly as follows:

Steps:

• Put the data files in .Rdata format in a directory containing only them. (The R objects contained in those files must be either SummarizedExperiment or ExpressionSet experiment.)
• Load all the experiments in a list using the function load_experiments

experiments <- load_experiments('directory_path')
experiments<-load_experiments('./')

If user want to download data from Expression Atlas, they can skip the above and go directly to "Downloading data from Expression Atlas".

Downloading data from Expression Atlas: In case of data downloading from Expression Atlas, user can use the function download_experiments_from_ExpressionAtlas in this way :

experiments <- download_experiments_from_ExpressionAtlas('E-MTAB-8549','E-MTAB-5060')

*** will download RNAseq experiments from expressionatlas.

Recommendation: Microarray input data should be preprocessed- appropriately back-ground corrected without any normalization step, probe-to-gene level mapped and log-transformed. It is also recommended to remove low-expressed genes from data if possible. For bulk-RNaseq data, it should be a count matrix.

Output from step 1: This downloads the experiments in a new directory called "experiments" in user's working directory and loads all the experiments in R within a list, using load_experiments function. After having loaded the experiments, user will get a list of either SummarizedExperiment or list of ExpressionSet objects.

Caution: Avoid mixing experiments of SummarizedExperiment with ExpressionSet. Experiments can only belong from any one of the classes only. All the selected experiments should have same gene id format.

Step2: Removing the isolated experiments

To correct the batch effect, one needs to take the biological characteristics of the samples into account (gender in our example). If no sample of an experiment shares biological characteristics with samples from other batches, it is not possible to correct batch effect with these batches because one cannot distinguish the biological difference from the artifact. The function remove_isolated_experiments removes the isolated experiments and plots graphs of intersections between the experiments before and after removal.

experiments %<>% remove_isolated_experiments('sex')

WARNING: this function only removes the isolated experiment according to the supplied key. Although it is still possible that two or more unconnected groups of experiments remain, within which the experiments are connected. In this case, batch effect correction is not possible and one will have to choose a group of experiments manually.

The two following plots are displayed by the function. The first one shows the graph of intersections of all the experiments before the removal of isolated ones. The second shows the same graph after their removal (Figure2).

Step3: Merging experiments in a single dataset

The function merge_experiments merges all the experiments in the list in a single SummarizedExperiment or ExpressionSet object and doesn't perform any correction. This function has two additional arguments log and filter (respectively set to TRUE and FALSE by default).

• For RNAseq experiments: users have to call merge_experiment.SummarizedExperiment() function.
• For microarray experiment, users have to call merge_experiment.ExpressionSet() function.
• The log argument determines whether to perform log transformation on the data (recommended for bulk RNAseq).
• The filter argument determines whether to filter genes for which all the samples of a batch have zero-counts. Set it to TRUE if user have issues in running ComBat at the next step.

experiments %<>% merge_experiment.SummarizedExperiment()

experiments is now an only ExpressionSet object containing the information about batches both in its @metadata and @colData slots :

Caution: Sometimes during merging experiment, phenodata (SDRF) file gets corrupted, hence, it is advised to always check meta-data before proceeding further.

Step4: Correcting batch effect

The function batch_correction performs various methods of batch-correction on a given merged dataset and output batch-corrected data as a list.

Short detail of methods implemented in batch_correction function are given below-

• For Microarray, batch_correction function has following batch correction methods-

  • limma- default limma setting, more detail about the method can be learnt from Limma documentation.
  • GFS- Gene Fuzzy Score, more detail about the method can be learnt from publication.
  • Robust quantile normalization- quantile normalisation method from PreprocessCore package. More information can be obtained from their documentation.
  • ComBat- implemented from SVA package. More information can be obtained from their documentation. In the current batch_correction method, there are two versions of ComBat- 1) ComBat1- for parametric adjustment and, 2) ComBat2 - for non-parametric adjustment, mean-only version.
  • Q_ComBat - is Quantile+ parametric adjustment of ComBat.
  • MNN - default mnnCorrect function from batchelor package. More information can be obtained from their documentation.
  • naiveRandRUV_HK - default naiveRandRUV using Human HK genes from RUVnormalize package. More information can be obtained from their documentation. The Human Housekeeping gene list is from this publication. No. of confounders were estimated using both "leek" and "be" method and then least no. of estimated confounders were used as input for all variants of naiveRandRUV method.
  • Q_naiveRandRUV_HK - is Qunatile+ naiveRandRUV_HK
  • naiveRandRUV_empi.controls - default naiveRandRUV using data-empirically derived HK genes from RUVnormalize package.
  • Q_naiveRandRUV_empi.controls -is Qunatile+ naiveRandRUV_empi.controls.

• For bulk-RNAseq, batch_correction function has the following batch correction methods-

  • limma- default limma setting, more detail about the method can be learnt from Limma documentation.
  • ComBat- implemented from SVA package. More information can be obtained from their documentation.
  • ComBat_seq - RNAseq version of ComBat, now implemented in SVA package. More information can be obtained from this publication.
  • MNN - default mnnCorrect function from batchelor package. More information can be obtained from their documentation.
  • RUVs - default RUVs function from RUVSeq package.Here, k=1 is used. More information can be obtained from their documentation.
  • scBatch - batch correction method implemented in scBatch package. More information can be obtained from their publication.

This function has following arguments-

• The experiment argument is the input merged dataset.
• The model argument is a R formula mentioning the biological factor to take into account during correction.
• The batch argument is a meta-data column which has information about batch-label.
• The k argument is a no. of confounders for RUV methods and denotes the "number of factors of unwanted variation to remove". If k is not provided it will then be estimated from input data.
• The filter argument specifies the gene label for the given dataset. It Should be one of the following string- 'symbol', 'ensembl_gene_id' or 'entrezgene_id' depending on the gene label for the given dataset.

batch effect present in the example study can be detected as follows:

dds_final <- DESeqDataSetFromMatrix(countData = assay(experiments  ),
                                    colData = colData(experiments), design = ~sex)

####check batch effect in the initial dataset:

vst3 <- varianceStabilizingTransformation(dds_final)

plotPCA(vst3,"batch")

dds_final <- estimateSizeFactors(dds_final)
se <- SummarizedExperiment(log2(counts(dds_final, normalized=TRUE) + 1),
                           colData=colData(experiments))
# the call to DESeqTransform() is needed to trigger our plotPCA method.
plotPCA( DESeqTransform( se ),intgroup =c("batch") )

Batch_correction using SelectBCM tool:

result.1 <- batch_correction(dds_final,model=~sex, batch = "batch")
result.1$data.uncorrected <- se
result.1$data.uncorrected1 <- vst3

Output: is the list of batch-corrected data:

summary(result.1)

                     Length Class                Mode
data.limma           31686  SummarizedExperiment S4  
data.ComBat          31686  SummarizedExperiment S4  
data.ComBat_Seq_full 31686  SummarizedExperiment S4  
data.ComBat_seq_null 31686  SummarizedExperiment S4  
data.MNN             31686  SummarizedExperiment S4  
data.RUVs            31686  SummarizedExperiment S4  
data.scBatch         31686  SummarizedExperiment S4  
data.uncorrected     31686  SummarizedExperiment S4  
data.uncorrected1    31686  DESeqTransform       S4 

Step5: Assessment of batch-correction methods

The batch_evaluation allows users to performs various evaluations on batch-corrected data and output performance list of each method. Since there are no. of ways batch-correction can be evaluated and each method has some limitation, we have used a cocktail of methods to perform analysis. This function has both PCA-inspired as well as biology inspired qualitative assessment protocol for batch-correction.

• For RNAseq experiments: users have to call batch_evaluation.RNAseq() function.
• For microarray experiment: users have to call batch_evaluation.microarray() function.

Short detail of methods implemented in batch_evaluation function is given below-

• pvca- A function for principal variance component analysis. The function is written based on the 'pvcaBatchAssess' function of the PVCA R package and slightly changed to make it more efficient and flexible for both microarray and bulk RNAseq gene-expression data. From https://github.com/dleelab/pvca.
• silhouette- Determine batch effect using the silhouette coefficient (adopted from scone) with default setting (nPcs=3). Taken from kBET package.
• pcRegression- Determines batch effect by a linear model fit of principal components and a batch (categorical) variable with a default setting (n_top=20). Taken from kBET package.
• entropy- Determines the batch effect by computing the entropy of mixing. It is a parameter to quantify the extent of the intermingling of cells from different batches. Taken from MNN package.For calculation, first two Pcs are used as input. Since,depending on no. of samples in merged dataset, it is important to choose N1 and N2 for Batchentropy calculation, we have put both arguments as variables here.
• HVG.union- This analysis calculates fraction of conserved highly variable genes (HVG) according to Brennecke et al., 2013 publication. This calculation also accounts for perseverance of biological heterogeneity by measuring ratio of conservation of HVG genes after batch correction. This function calculates ratio of number of conserved HVG after batch correction/ Union of HVG genes among all the batches. Therefore, this function is less stringent compared to HVG.intersection function.

This function has following arguments-

• The result argument is a list of wrapped batch-corrected experiments obtained from the last step ('batch_correction').
• The batch.factors is a list of factors to perform PVCA analysis. Along with batch as factor, one biological factor which can be used to assess over-fitting should be provided. Providing more than one biological factor here, will create issues while plotting the results in next step.
• The experiment argument is the input merged dataset.
• The N1 is the number of randomly picked samples for the BatchEntropy function.
• The N2 is the number of nearest neighbours of the sample (from all batches) to check (for BatchEntropy function).
• The filter argument specifies the gene label for the given dataset. It Should be one of following string- 'symbol', 'ensembl_gene_id' or 'entrezgene_id' depending on the gene label for the given dataset.

Assessment of batch-corrected data can be performed as follows:

assess.RNAseq <- batch_evaluation.RNAseq(result.1, batch.factors=c("batch","sex"),
                                         se,5,5,'ensembl_gene_id')

Output: assess.RNAseq is a nested list of evaluation scores for each of the evaluation protocols.

Step6: Ranking of BCMs

Once assessment is performed, in next step, results obtained from the batch_evaluation step, ranking of BCM is the two step analysis-

• Generation of Diagnostic matrix
• Plot of Ranks of BCM obatined from Diagnostic matrix

Diagnostic matrix

diagnostic_matrix.RNAseq function performs ranking of evaluation data obtained at the previous step. Here, methods are ranked based on their performance and finally sumRank is the final Rank of each method for the given input dataset. Rank 1 will be the best performer method.

• For RNAseq experiments: users have to call diagnostic_matrix.RNAseq function.
• For microarray experiment: users have to call diagnostic_matrix.microarray function.

This function has only one argument-

• diagnostic is an evaluation list obtained from the previous step batch_evaluation.

final <- SelectBCM::diagnostic_matrix.RNAseq(assess.RNAseq)

Output1: A list of two data-frame- (a) raw - simple data-frame output of evaluation matrix and, (b) ranked- Ranked data-frame of evaluation matrix which has additional column sumRank containing final Rank of each method. Ranks are in descending performance order, i.e. the method having score 1 will be the best method.

diagnostic_plot

Using output1, selectBCM provides-

a) Diagnostic plot showing the performance of BCMs across evaluation methods, where x-axis is the evaluation protocol and y-axis is the Rank of each batch-correction method. b) violin plot to summarise the performance of BCMs

• For RNAseq experiments: users have to call diagnostic_plot.RNAseq function.
• For microarray experiment: users have to call diagnostic_plot.microarray function.

dia_plot <-  SelectBCM::diagnostic_plot.RNAseq(final)

svg("Diagnostic_plot_RNAseq_example.svg")

ggarrange(dia_plot$diagnostic_plot,dia_plot$diagnostic_stat_summary,
          labels = c("a.", "b."),
          legend = "right", ncol = 1, nrow = 2
)

dev.off()

Rank plot

A bar plot representation of the BCMs where the 1st rank represents the most appropriate BCM for the given set of experiments using sumRank score.

bcm_ranking(final)

Additional function: Detetction of batch-effects in raw merged dataset

detect_effect performs a subset of evaluation tests of batch_evaluation function and provides estimates whether the merged dataset obtained after 'merge_experiments' requires batch correction or not. Higher value of pvca.batch, silhouette, pcRegression and entropy is a direct indicative of batch-effects in raw merged dataset.

This function has following arguments-

• The result argument is a is the merged dataset obtained from merge_experiments step.
• The batch.factors is a list of factors to perform PVCA analysis. Along with batch as factor, one biological factor which can be used to assess over-fitting should be provided. Providing more than one biological factor here, will create issues while plotting the results.
• The experiment argument is again the same merged dataset.
• The N1 is the number of randomly picked samples for the BatchEntropy function.
• The N2 is the number of nearest neighbours of the sample (from all batches) to check (for BatchEntropy function).
• The filter argument specifies the gene label for the given dataset. It Should be one of the following string- 'symbol', 'ensembl_gene_id' or 'entrezgene_id' depending on the gene label for the given dataset.

Assessment of batch-effect on raw merged data can be performed like this:

batch_effect.raw <- detect_effect(experiments,experiment = experiments,
                                  batch.factors=c("batch","sex"),5,5,'ensembl_gene_id')
Result <- do.call(rbind, lapply(batch_effect.raw, as.matrix)) %>% t %>% as.data.frame()

Result:

Result

       batch sex silhouette pcRegression Entropy
raw 0.6578993   0  0.7302323    0.8403614       0