In mathewchamberlain/SignacX: Cell Type Identification and Discovery from Single Cell Gene Expression Data

Sometimes we don't have time to run Signac and need a faster solution. Although Signac scales fine with large data sets (>300,000 cells) and even for large data, typically takes less than an hour, we developed SignacFast to quickly classify single cell data. Unlike Signac, SignacFast uses a pre-trained ensemble of neural network models generated from the HPCA reference data, speeding classification time ~5-10x fold. These models were generated from the HPCA training data like so:

# load pre-trained neural network ensemble model
ref = GetTrainingData_HPCA()

# generate models
Models_HPCA = ModelGenerator(R = training_HPCA, N = 100, num.cores = 4)

The "Models_HPCA" are accessed from within the R package:

# load pre-trained neural network ensemble model
Models = GetModels_HPCA()

We demonstrate how to use SignacFast in this vignette, which shows that SignacFast is broadly consistent with Signac (just faster). Here, we show how to use SignacFast to annotate flow-sorted synovial cells by integrating SignacX with Seurat. We start with raw counts from this publication.

all_times <- list()  # store the time for each chunk
knitr::knit_hooks$set(time_it = local({
  now <- NULL
  function(before, options) {
    if (before) {
      now <<- Sys.time()
    } else {
      res <- difftime(Sys.time(), now, units = "secs")
      all_times[[options$label]] <<- res
    }
  }
}))
knitr::opts_chunk$set(
  tidy = TRUE,
  tidy.opts = list(width.cutoff = 95),
  message = FALSE,
  warning = FALSE,
  time_it = TRUE
)
#celltypes_fast = readRDS("./fls/celltypes_fast_citeseq.rds")
#celltypes = readRDS("./fls/celltypes_citeseq.rds")
# pbmc = readRDS("fls/pbmcs_signac_citeseq.rds")
celltypes = readRDS(file = "fls/celltypes_amp_synovium.rds")
celltypes_fast = readRDS(file = "fls/celltypes_fast_synovium_celltypes.rds")

Load data

Read the CEL-seq2 data.

ReadCelseq <- function (counts.file, meta.file)
{
  E = suppressWarnings(readr::read_tsv(counts.file));
  gns <- E$gene;
  E = E[,-1]
  E = Matrix::Matrix(as.matrix(E), sparse = TRUE)
  rownames(E) <- gns
  E
}

counts.file = "./fls/celseq_matrix_ru10_molecules.tsv.gz"
meta.file = "./fls/celseq_meta.immport.723957.tsv"

E = ReadCelseq(counts.file = counts.file, meta.file = meta.file)
M = suppressWarnings(readr::read_tsv(meta.file))

# filter data based on depth and number of genes detected
kmu = Matrix::colSums(E != 0)
kmu2 = Matrix::colSums(E)
E = E[,kmu > 200 & kmu2 > 500]

# filter by mitochondrial percentage
logik = grepl("^MT-", rownames(E))
MitoFrac = Matrix::colSums(E[logik,]) / Matrix::colSums(E) * 100
E = E[,MitoFrac < 20]

Seurat

Start with the standard pre-processing steps for a Seurat object.

library(Seurat)

Create a Seurat object, and then perform SCTransform normalization. Note:

• You can use the legacy functions here (i.e., NormalizeData, ScaleData, etc.), use SCTransform or any other normalization method (including no normalization). We did not notice a significant difference in cell type annotations with different normalization methods.
• We think that it is best practice to use SCTransform, but it is not a necessary step. Signac will work fine without it.

# load data
synovium <- CreateSeuratObject(counts = E, project = "FACs")

# run sctransform
synovium <- SCTransform(synovium)

Perform dimensionality reduction by PCA and UMAP embedding. Note:

• Signac actually needs these functions since it uses the nearest neighbor graph generated by Seurat.

# These are now standard steps in the Seurat workflow for visualization and clustering
synovium <- RunPCA(synovium, verbose = FALSE)
synovium <- RunUMAP(synovium, dims = 1:30, verbose = FALSE)
synovium <- FindNeighbors(synovium, dims = 1:30, verbose = FALSE)

SignacX

library(SignacX)

Generate Signac labels for the Seurat object. Note:

• Optionally, you can do parallel computing by setting num.cores > 1 in the Signac function.

labels <- Signac(synovium, num.cores = 4)
celltypes = GenerateLabels(labels, E = synovium)

Sometimes, training the neural networks takes a lot of time. The above classification took 27 minutes. To make a faster method, we implemented SignacFast which uses pre-trained models. Note:

• SignacFast uses an ensemble of 1,800 pre-calculated neural networks using the GenerateModels function together with the training_HPCA reference data set.
• Features that are absent from the single cell data and present in the neural network are set to zero.

# Run SignacFast
labels_fast <- SignacFast(synovium, num.cores = 4)
celltypes_fast = GenerateLabels(labels_fast, E = synovium)

Compare results:

Celltypes:

knitr::kable(table(Signac = celltypes$CellTypes, SignacFast = celltypes_fast$CellTypes), format = "html")

Cellstates:

knitr::kable(table(Signac = celltypes$CellStates, SignacFast = celltypes_fast$CellStates), format = "html")

Save results

saveRDS(synovium, file = "fls/seurat_obj_amp_synovium.rds")
saveRDS(celltypes, file = "fls/celltypes_amp_synovium.rds")
saveRDS(celltypes_fast, file = "fls/celltypes_fast_amp_synovium_celltypes.rds")

write.csv(x = t(as.data.frame(all_times)), file = "fls/tutorial_times_SignacFast_AMP.csv")

Session Info

sessionInfo()

mathewchamberlain/SignacX documentation built on March 3, 2023, 2:46 a.m.