FastIntegration v1.1.0

FastIntegration is a fast and high-capacity version of Seurat Integration. FastIntegrate can integrate thousands of scRNA-seq datasets and outputs batch-corrected values for downstream analysis.

Recent update: New functions which allow users to filter and download data in DISCO (https://www.immunesinglecell.org/repository), comprising 5200+ single-cell samples!

More vignettes can be found at https://immunesinglecell.org/vignette/docs/FastIntegration/cell-type-atlas

Requirement

FastIntegration requires the following packages:

• R (>= 4.0.0)
• Seurat (>= 4.0.0)
• SeuratObject (>= 4.0.0)
• data.table
• Matrix
• tictoc
• dplyr
• pbmcapply

We highly recommend you to build R with openblas which will accelerate integration 2-3x times.

Here is the common way to do it:

sudo yum install -y openblas openblas-threads openblas-openmp # for centos

sudo apt-get install libopenblas-dev # for debian

./configure --enable-R-shlib --enable-byte-compiled-packages --enable-BLAS-shlib --enable-memory-profiling

--with-blas="-lopenblas"

Installation

devtools::install_github("git@github.com:JinmiaoChenLab/FastIntegrate.git")

Usage

Preprocess

library(Seurat)
library(pbmcapply)
rna.list = readRDS("rna_list.rds") # read list of Seurat object, each element in list is a sample

# make all samples have same genes
overlapped.gene = Reduce(intersect, lapply(rna.list, rownames))
for (i in 1:length(rna.list)) {
  rna.list[[i]] = subset(rna.list[[i]], features = overlapped.gene)
  rna.list[[i]] = NormalizeData(rna.list[[i]])
  rna.list[[i]] = FindVariableFeatures(rna.list[[i]])
  rna.list[[i]] = RenameCells(rna.list[[i]], new.names = paste0(Cells(rna.list[[i]]), "--", i))
}

Onestop function

For large sample size (> 200 samples), we recommend to use step by step integration.

library(FastIntegration)
# rna.list is the list of seurat object
data = OneStopIntegration(
  rna.list = rna.list, 
  tmp.dir = "./test/", 
  max.cores = 30
)

Step by step integration

library(Seurat)
library(pbmcapply)
library(FastIntegration)

# rna.list is the list of seurat object
BuildIntegrationFile(rna.list = rna.list, tmp.dir = "./", nCores = 50)
FastFindAnchors(tmp.dir = "./", nCores = 50)

# close current R session and open a new one to clean the memory (This is important for large data integration)
# In the new session, please just set work directory and do not load any data. Then run the following codes:

library(Seurat)
library(pbmcapply)
library(FastIntegration)

genes = readRDS("FastIntegrationTmp/raw/1.rds")
genes = rownames(genes)
idx = split(1:length(genes), cut(1:length(genes), 20, labels = FALSE))
pbmclapply(
  1:20, function(i) {
    rna.integrated = FastIntegration(tmp.dir = "./", npcs = 1:30, slot = "data",
                                     features.to.integrate = genes[idx[[i]]])
    saveRDS(rna.integrated, paste0("FastIntegrationTmp/inte/inte_", i, ".rds"), compress = F)
  }, mc.cores = 20 
)

After integration

##### create Seurat obj with the variable features of integration (For very big dataset) ##### 
features = readRDS("FastIntegrationTmp/others/features.rds")
rna.data = pbmclapply(
  1:20, function(i) {
    rna = readRDS(paste0("./FastIntegrationTmp/inte/inte_", i, ".rds"))
    rna = rna[intersect(rownames(rna), features),]
    return(rna)
  }, mc.cores = 20
)
rna.data = do.call(rbind, rna.data)
rna.data = CreateSeuratObject(rna.data)
rna.data = ScaleData(rna.data, features = features)
rna.data = RunPCA(rna.data, features = features, npcs = 50)
rna.data = FindNeighbors(rna.data, dims = 1:50)
rna.data = FindClusters(rna.data, graph.name = "RNA_snn", algorithm = 2)
rna.data = RunUMAP(rna.data, dims = 1:50)


##### select varibale gene based on integrated data  (For dataset with less than 100 samples) #####
rna.data = pbmclapply(
  1:20, function(i) {
    rna = readRDS(paste0("FastIntegrationTmp/inte/inte_", i, ".rds"))
    return(rna)
  }, mc.cores = 20
)

rna.data = do.call(rbind, rna.data)
rna.data = CreateSeuratObject(rna.data)
rna.data = FindVariableFeatures(rna.data, nfeatures = 2000)
features = VariableFeatures(rna.data)
rna.data = ScaleData(rna.data, features = features)
rna.data = RunPCA(rna.data, features = features)
rna.data = FindNeighbors(rna.data, dims = 1:50)
rna.data = FindClusters(rna.data, resolution = 0.5, algorithm = 2)
rna.data = RunUMAP(rna.data, dims = 1:50)

Download data from DISCO

##### Filter samples and get metadata ##### 
# You can filter samples by their different headers: tissue, disease, platform, project.id. sample.id, sample.type
# For each header, you can select multiple items as follows:
meta = FindSampleByMetadata(tissue = c("blood", "kidney"))

##### Download sample ##### 
# You can further filter cells by specifying expressed or unexpressed genes.
# dir is the location where the files are saved
DownloadDiscoData(meta, expressed.gene = c("CD3E"), unexpressed.gene = c("CD8A"), dir = "./disco") # mostly CD4 T cells (CD3E+CD8A-)

Usage Scenario

We have apply FastIntegration to DISCO database for integrating thousands of samples.

License

All other code in this repository is licensed under a GPL-3 license.