[go: up one dir, main page]

Skip to content

casanova-lab/iMUBAC

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

39 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

iMUBAC: Integration of Multi-Batch Cytometry Datasets

The 'iMUBAC' package provides a structured framework for objective inter-batch comparisons and unbiased immunophenotyping of high-dimentional cytometry datasets.

System Requirements

iMUBAC has been tested on R versions >= 4.0 on Windows platform.

Installation

Install the latest version as follows:

if(!require(devtools)) install.packages("devtools")
devtools::install_github("VPetukhov/ggrastr")
devtools::install_github("immunogenomics/harmony")
devtools::install_github("casanova-lab/iMUBAC")
devtools::install_github("masato-ogishi/plotUtility")
  •  You may be prompted to install some packages before installing iMUBAC. Follow the messages.
  •  This package depends on some packages in the Bioconductor (e.g., flowCore) that may not be automatically installed. Please check the DESCRIPTION file for more details on required packages and install them manually as required.
  • You need an appropriate rJava setting beforehand.

Usage

  1. Working environment
options(java.parameters="-Xmx60G")  ## allow JAVA to use large memory space
library(tidyverse)
library(data.table)
library(SingleCellExperiment)
library(iMUBAC)
  1. Data
  •  For demonstration, previously published CyTOF datasets from Krieg et al. (Pre-treatment datasets stained with the Panel 3, a myeloid-specific panel, from two different batches) are included.
# Sample metadata
md <- data.table::fread(system.file("Metadata.csv", package="iMUBAC"))
md[,full_path:=file.path(system.file(package="iMUBAC"),file_name)]
md[,batch:=factor(batch, levels=c("Data23","Data29"))]
md[,group:=factor(group, levels=c("HD","R","NR"))]

# Panel design
pd <- list(
  data.table::fread(system.file("PanelInfo_Data23_Panel3.csv", package="iMUBAC")),
  data.table::fread(system.file("PanelInfo_Data29_Panel3.csv", package="iMUBAC"))
) %>% magrittr::set_names(c("Data23","Data29"))

# Preprocessing
sce <- iMUBAC::prepSCE(
  md=md[panel=="Panel3"],
  pd=pd,
  channel_length=NULL,      ## Doublet detection by event length is disabled.
  channel_DNA=c("Ir191Di","Ir193Di"),
  channel_LD="Pt198Di",
  type="CyTOF"
)
colData(sce) <- colData(sce)[c("file_name","panel","batch","donor_id","group","treatment")]
saveRDS(sce, "CyTOF_SCE_Merged_Panel3.rds")
  1. Batch-correction
  •  Here, only control samples (in this case, "HD") are used for batch-correction. However, this is the users' call.
  • The harmonization step takes some time, depending on the number of batches and cells per batch to be integrated.
# Load the preprocessed data
sce <- readRDS("CyTOF_SCE_Merged_Panel3.rds")

# Take the subset of data from healthy donors
sce_down <- sce[,sce$"group"=="HD"]

# Batch-correction
sce_down <- iMUBAC::batchCorrection(
  sce_down,
  maxN=50000, ## A maximum of 50000 cells are randomly selected from each batch.
  seed=12345  ## a random seed
)
  •  Optionally, we can explore the impact of batch-correction through UMAP.
# UMAP
sce_down <- iMUBAC::runUMAP(
  sce_down,
  by_exprs_values="exprs",
  name="UMAP",
  n_neighbors=25,
  min_dist=0.4,
  scale=T,
  n_threads=4, ## the number of threads for parallel computing
  seed=12345
)
sce_down <- iMUBAC::runUMAP(
  sce_down,
  by_exprs_values="normexprs",
  name="UMAPnorm",
  n_neighbors=25,
  min_dist=0.4,
  scale=T,
  n_threads=4,
  seed=12345
)

# Plot
plt1 <- ggpubr::ggarrange(
  iMUBAC::plotDR(
    sce_down,
    dimred="UMAP",
    colour_by="batch"
  ) +
    scale_colour_brewer(palette="Dark2"),
  iMUBAC::plotDR(
    sce_down,
    dimred="UMAPnorm",
    colour_by="batch"
  ) +
    scale_colour_brewer(palette="Dark2"),
  ncol=2, nrow=1, common.legend=T, legend="right"
)
plotUtility::savePDF(plt1, o="Plt1.pdf", w=15, h=6)
  1. Unsupervised clustering
  •  Currently two methods are implemented.
  • The first approach groups cells into \code{xdim}x\code{ydim} clusters using \pkg{FlowSOM}, and then performs metaclustering with \pkg{ConsensusClusterPlus} into \code{maxK} clusters.
  • The second approach performs dimention reduction through UMAP, constructs shared nearest-neighbor graphs using \pkg{scran}, and then performs community detection using the Leuvein algorithm in \pkg{igraph}.
# FlowSOM-guided clustering & ConsensusClusterPlus-guided metaclustering
sce_down <- iMUBAC::clustering(
  sce_down,
  features=rownames(sce_down), ## Using all markers for clustering
  by_exprs_values="normexprs",
  method="FlowSOM",
  xdim=20,
  ydim=20,
  maxK=40, ## the number of metaclusters returned
  seed=12345
)

# Alternatively, SNN-graph-guided clustering
## Not used in this demonstration
sce_down_snn <- iMUBAC::clustering(
  sce_down,
  features=rownames(sce_down), ## Using all markers for clustering
  by_exprs_values="normexprs",
  method="SNNGraph",
  n_components=10, ## the number of reduced dimentions for constructing the SNN graph
  n_neighbors=25,  ## the parameters for UMAP-based dimention reduction
  min_dist=0.4,    ## the parameters for UMAP-based dimention reduction
  seed=12345
)
  •  We can explore the clusters through the median expression heatmap and UMAP plots.
# Median expression heatmap
plt2 <- iMUBAC::plotClusterHeatmap(
  sce_down,
  features=rownames(sce_down),
  clusters=sce_down$"cluster_id",
  by_exprs_values="normexprs",
  fun="median",
  scale=T,
  cluster_rows=T,
  cluster_anno=T,
  draw_dend=T,
  draw_freqs=T
)
plotUtility::savePDF(plt2, o="Plt2.pdf", w=12, h=8)

# UMAP plot
plt3 <- iMUBAC::plotDR(
  sce_down,
  dimred="UMAPnorm",
  colour_by="cluster_id",
  text_by="cluster_id"    ## to overlay cluster ids on each of the clusters
) +
  ggpubr::rremove("legend")
plotUtility::savePDF(plt3, o="Plt3.pdf", w=10, h=6)
  1. Batch-specific cluster propagation through machine learning
  •  Classifiers are trained for each batch utilizing the Extreme Randomized Trees algorithm.
  • There are both abundant and rare clusters. To mitigate the class imbalance problem and also to reduce the computational burden, here a maximum of 100 cells are randomly selected for each of the clusters before classifier training.
sce <- iMUBAC::clusterPropagation(
  sce,           ## the original data containing cells from controls and patients (non-batch-corrected)
  sce_down,      ## down-sampled data containing cells from controls (batch-corrected and clustered)
  by_exprs_values="exprs", ## Non-batch-corrected expression values are used for classifier training.
  maxN=100,
  numThreads=4,  ## the number of threads for parallel computing
  seed=12345
)
  1. Manual cell-type identification
  •  Users can provide cell-type annotations for each cluster. The data.frame should contain four columns: "cluster_id", "celltype", "celltype_detailed", and "order".
  • The "celltype" and "celltype_detailed" columns can be identical but may be useful to provide different layers of annotation (e.g., "CD4 T" and "CD4 TEMRA").
  • The "order" column will be used to define the order of factor levels for cell types. Expected to be filled with integers.
# Annotation
df_celltype <- readxl::read_excel("CyTOF_ClusterCellTypes_Panel3.xlsx")
sce <- clusterAnnotation(sce, df_celltype)
sce_down <- clusterAnnotation(sce_down, df_celltype)
saveRDS(sce, "CyTOF_SCE_Cluster_Panel3.rds")
saveRDS(sce_down, "CyTOF_SCE_Cluster_Panel3_Down.rds")

# Median expression heatmap
## Removing unnecessary clusters
sce_down <- sce_down[,!sce_down$"celltype_detailed" %in% c("Basophil","Unidentified")]
sce_down$"celltype_detailed" <- droplevels(sce_down$"celltype_detailed")
plt4 <- iMUBAC::plotClusterHeatmap(
  sce_down,
  features=rownames(sce_down),
  clusters=sce_down$"celltype_detailed",
  by_exprs_values="normexprs",
  fun="median",
  scale=T,
  cluster_rows=F,
  cluster_anno=F,
  draw_dend=T,
  draw_freqs=T
)

# UMAP
## Recompute UMAP coordinates after removing unnecessary clusters
sce_down <- sce_down[,!sce_down$"celltype_detailed" %in% c("Basophil","Unidentified")]
sce_down$"celltype_detailed" <- droplevels(sce_down$"celltype_detailed")
sce_down <- iMUBAC::runUMAP(
  sce_down,
  by_exprs_values="normexprs",
  name="UMAP",
  n_neighbors=25,
  min_dist=0.4,
  scale=T,
  n_threads=4,
  seed=12345
)
plt5 <- iMUBAC::plotDR(
  sce_down,
  dimred="UMAP",
  colour_by="celltype_detailed",
  text_by="celltype_detailed",
  point_size=1.5
) +
  theme(legend.position="right",
        legend.direction="vertical") +
  ggpubr::rremove("legend.title")
  1. Differential abundance analysis
  • A number of workflows are available for differential abundance (DA) analysis. Here, we use the QLF test in edgeR to detect DA subsets between non-responders and responders to PD-1 blockade immunotherapy.
library(edgeR)
sce <- readRDS("CyTOF_SCE_Cluster_Panel3.rds")
dt_da <- colData(sce) %>%
  as.data.frame() %>%
  dplyr::group_by(donor_id, batch, group, celltype_detailed) %>%
  dplyr::tally() %>%
  dplyr::mutate(percent=n/sum(n)*100) %>%
  tidyr::complete(tidyr::nesting(donor_id, batch, group), celltype_detailed, fill=list(n=0, percent=0)) %>%
  data.table::as.data.table()
dt_da[,batch:=factor(batch, levels=c("Data23","Data29"), labels=c("Batch1","Batch2"))]
dt_da[,variable:=paste0(group,"_",batch)]
dt_da[,variable:=factor(variable, levels=c("HD_Batch1", "HD_Batch2", "R_Batch1", "R_Batch2", "NR_Batch1", "NR_Batch2"))]
dt_da[,label:=paste0(group,"_",batch,"_",donor_id)]
dt_da_wide <- data.table::dcast(dt_da, celltype_detailed~label, value.var="n")
celltypeLabels <- dt_da_wide$"celltype_detailed"
dt_da_wide[,celltype_detailed:=NULL]
mat <- as.matrix(dt_da_wide)
rownames(mat) <- celltypeLabels
groups <- factor(
  colnames(mat),
  levels=unique(dt_da, by="label")$"label",
  labels=unique(dt_da, by="label")$"group"
)
batches <- factor(
  colnames(mat),
  levels=unique(dt_da, by="label")$"label",
  labels=unique(dt_da, by="label")$"batch"
)
d <- edgeR::DGEList(counts=mat, lib.size=colSums(mat), group=groups, remove.zeros=T)
design <- model.matrix(~0+groups+batches, data=d$samples)
d <- edgeR::estimateDisp(d, design)
fit <- edgeR::glmQLFit(d, design, robust=T)
dt_da_qlf <- edgeR::glmQLFTest(
  fit, 
  contrast=makeContrasts(NRvsR=groupsNR-groupsR, levels=design)
) %>%
  edgeR::topTags(n=Inf) %>%
  as.data.frame() %>%
  dplyr::transmute(
    celltype_detailed=rownames(.),
    log2FC=logFC, 
    PValue=PValue,
    AdjPValue=FDR
  ) %>%
  data.table::as.data.table()

Reference

Ogishi, M et al. (2020) "Multibatch Cytometry Data Integration for Optimal Immunophenotyping." The Journal of Immunology. https://www.jimmunol.org/content/early/2020/11/20/jimmunol.2000854
Krieg, C et al. (2018) "High-dimensional Single-Cell Analysis Predicts Response to anti-PD-1 Immunotherapy." Nature Medicine. https://pubmed.ncbi.nlm.nih.gov/29309059/

Releases

No releases published

Packages

No packages published

Languages