Introduction
Clustering is the key step to define cell types from the heterogeneous cell population. One critical challenge is that single-cell RNA sequencing (scRNA-seq) experiments generate a massive amount of data with an excessive noise level, which hinders many computational algorithms. We propose a single cell Clustering method using Autoencoder and Network fusion (scCAN) that can accurately segregate cells from the high-dimensional noisy scRNA-Seq data.
Preliminaries
This vignette provides step-by-step instruction for scCAN R package installation, analyses using example and scRNA-seq datasets, and an in-depth analysis using scRNA-seq data with prior biological knowledge.
Install scCAN package from CRAN repository.
install.packages("scCAN")To start the analysis, load scCAN package:
library(scCAN)## Loading required package: scDHA## Loading required package: FNN## Loading required package: purrrAnalysis on the example dataset
The input of scCAN is an expression matrix in which rows represent samples and columns represent genes. scCAN package includes an example dataset of 1,000 cells and 2,000 genes.
#Load example data (SCE dataset)
data("SCE")
#Get data matrix and label
data <- t(SCE$data); label <- as.character(SCE$cell_type1)Inspect the example of 10 cells and 10 genes.
data[1:10,1:10]## [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
## Cell1 41 57 45 12 0 0 6 6 1 129
## Cell2 73 55 45 3 8 0 30 33 0 176
## Cell3 46 62 40 3 11 0 5 21 0 106
## Cell4 37 39 13 0 16 0 5 0 0 61
## Cell5 77 76 72 0 18 0 10 39 4 140
## Cell6 23 34 16 0 0 0 5 21 0 62
## Cell7 0 15 11 0 0 0 6 10 0 31
## Cell8 58 56 63 8 9 0 4 20 0 112
## Cell9 24 20 34 1 0 0 2 38 0 46
## Cell10 71 61 61 8 14 0 20 23 0 0dim(data)## [1] 1000 2000Next, we can cluster the example dataset using the following command.
#Generate clustering result. The input matrix has rows as samples and columns as genes
result <- scCAN(data)
#The clustering result can be found here
cluster <- result$clusterUsers can visualize the expression data with the new cluster annotation. To reduce the dimension, users can use the package irlba [1] to obtain the first 20 principal components, and the package Rtsne [2] to obtain the transcriptome landscape using t-SNE. Finally, we can visualize the cell landscape using scatter plot available in ggplot2 [3] package.
suppressPackageStartupMessages({
library(irlba)
library(ggplot2)
library(Rtsne)
})
# Get 2D emebedding data of the original data.
set.seed(1)
low <- irlba::irlba(data, nv = 20)$u
tsne <- as.data.frame(Rtsne::Rtsne(low)$Y)
tsne$cluster <- as.character(cluster)
colnames(tsne) <- c("t-SNE1","t-SNE2","Cluster")
p <- ggplot2::ggplot(tsne, aes(x = `t-SNE1`, y = `t-SNE2`, colour = `Cluster`))+
ggtitle("Transcriptome landscape visualization using t-SNE")+labs(color='Cluster') +
geom_point(size=2, alpha = 0.8) +
theme_classic()+
theme(axis.text.x=element_blank(),
axis.ticks.x=element_blank(),
axis.text.y=element_blank(),
axis.ticks.y=element_blank(),
axis.title=element_text(size=14),
plot.title = element_text(size=16),
panel.grid.major = element_blank(),
panel.grid.minor = element_blank(),
legend.position="bottom", legend.box = "horizontal",
legend.text =element_text(size=10),
legend.title=element_text(size=14))+
guides(colour=guide_legend(nrow = 1))
p
Visualization of expression data using cells clusters identified from scCAN.
Download and analyze real human dataset
Data downloading and clustering
We will use a real dataset from Pollen et al.[4] that is available on our website. This dataset has 301 cells sequenced from human tissues including blood, skin, brain, and stem cells.
suppressPackageStartupMessages({
library(SummarizedExperiment)
})
path <- "https://bioinformatics.cse.unr.edu/software/scCAN/data/pollen.rds"
SCE <- readRDS(url(path,"rb"))
# Get expression matrix
data <- t(SummarizedExperiment::assay(SCE))
# Get cell annotation
label <- SummarizedExperiment::colData(SCE)Next, we can cluster dataset using the following command.
#Generate clustering result, the input matrix has rows as samples and columns as genes
start <- Sys.time()
result <- scCAN(data)
running_time <- difftime(Sys.time(),start,units = "mins")
print(paste0("Running time = ", running_time))
#The clustering result can be found here
cluster <- result$cluster## Running time = 1.293936 minsThis dataset also has an annotation file that includes true cell labels:
head(label)DataFrame with 6 rows and 2 columns
cell_type1 cell_type2
<factor> <factor>
Hi_2338_1 2338 dermal
Hi_2338_2 2338 dermal
Hi_2338_3 2338 dermal
Hi_2338_4 2338 dermal
Hi_2338_5 2338 dermal
Hi_2338_6 2338 dermalThe annotation file is stored data frame in which rows are samples. The annotation file has two sets of labels: (i) cell_type1 indicates cell types discovered from Pollen et al., (ii) cell_type2 has annotation of tissues that the cells was taken. Users can compare the obtain clustering results against cell_type1 for validation and visualization.
#Calculate adjusted Rand Index using mclust package
ari <- round(scCAN::adjustedRandIndex(cluster,label$cell_type1), 2)
print(paste0("ARI = ", ari))Users can use the same code above to plot the transcriptome landscape of the Pollen dataset. Below is the transcriptome landscapes of the dataset using the annotation provided by Pollen et al., and the landscape using the new cluster annotation.
Visualization of expression data using original cell types and cluster identified from scCAN for Pollen dataset.
scCAN was able to cluster Pollen dataset in 1 minutes with 9 clusters identified.The clusters discovered by scCAN are highly correlated with cell labels obtained from the original publication [4] with an ARI of 0.92.
Visualizing the data using latent variables of scCAN
scCAN generates a compressed representation of original expression data. This representation is part of the output, and can be obtained using the following code:
suppressPackageStartupMessages({
library(ggplot2)
library(cowplot)
library(scatterplot3d)
library(scales)
})
# Get latent data from scCAN result
latent <- result$latent
# Generating 2D embedding data using 2 first latent variables
data1= latent[,1:2]
# Generating 3D embedding data using 3 first latent variables
data2= latent[,1:3]
#Plotting 2D data using ggplot
name ="Pollen"
dat <- as.data.frame(data1)
dat$p <- as.character(cluster)
colnames(dat) <- c("Latent Variable 1", "Latent Variable 2", "scCAN Cluster")
p1 <- ggplot(dat, aes(x = `Latent Variable 1`,y = `Latent Variable 2`,colour = `scCAN Cluster`)) +
theme(plot.title = element_text(hjust = 0.5)) +
geom_point(size=2, alpha = 0.6) +
#scale_color_manual(values = colors)+
theme_classic()+
theme(axis.text.x=element_blank(),
axis.ticks.x=element_blank(),
axis.text.y=element_blank(),
axis.ticks.y=element_blank(),
axis.title=element_text(size=14),
panel.grid.major = element_blank(),
panel.grid.minor = element_blank(),
legend.position="none", legend.box = "horizontal",
legend.text =element_text(size=12),
legend.title=element_text(size=14) )+
guides(colour=guide_legend(ncol=2))
title1 <- ggdraw() + draw_label(name, size=16)
legend1 <- get_legend(p1 +
theme(legend.position="bottom", legend.box = "horizontal",legend.title=element_text(size=12),
legend.margin = margin(0), legend.spacing.x = unit(0.1, "cm") ,
legend.text=element_text(size=12)) + guides(color=guide_legend(nrow = 2))
)
m1 <- plot_grid(title1, p1, legend1, rel_heights = c(0.05, 0.5, .1), ncol = 1)
plot_grid(m1,rel_heights = c(0.05, 0.9), ncol = 1)
#Plotting 3D data using scatterplot3d package
graphics.off()
name = "Pollen"
dat <- as.data.frame(data2)
# dat$t <- label$cell_type1
dat$p <- as.character(cluster)
colnames(dat) <- c("Latent Variable 1", "Latent Variable 2", "Latent Variable 3", "scCAN Cluster")
pal <- hue_pal()(length(unique(cluster)))
colors <- pal[as.numeric(dat$`scCAN Cluster`)]
scatterplot3d(dat[,1:3], pch = 16,
grid=TRUE, box=TRUE,color = colors)The 2D and 3D visualizations using 2 latent variables and 3 latent variables obtained from latent data are shown as follows:
Visualization of 2D and 3D representation of scCAN latent data using clusters identified from scCAN for Pollen dataset.
Users can also use latent data as an input to standard dimensional reduction algorithms including built-in R Principal Component Analysis (PCA), t-SNE [5] and UMAP [6].
suppressPackageStartupMessages({
library(umap)
})
set.seed(1)
pc.data <- prcomp(latent, rank. = 2)$x
tsne.data <- Rtsne(latent, check_duplicates = F,dims = 2)$Y
umap.data <- umap(latent,n_components = 2)$layoutUsers can use the ggplot function to plot the 2D visualizations of the latent data:
2D Visualization of latent data with PCA, t-SNE and UMAP for Pollen dataset.
We can also use PCA, t-SNE and UMAP to get 3D embedding data of the latent variables.
suppressPackageStartupMessages({
library(Rtsne)
library(umap)
})
# Get latent data from scCAN result
latent <- result$latent
set.seed(1)
pc.data <- prcomp(latent, rank. = 3)$x
tsne.data <- Rtsne(latent, check_duplicates = F,dims = 3)$Y
umap.data <- umap(latent,n_components = 3)$layoutWe can use scatterplot3d R package to plot 3D visualization of the latent data.
suppressPackageStartupMessages({
library(ggplot2)
library(cowplot)
library(scatterplot3d)
library(scales)
})
graphics.off()
name = "Pollen"
par(mfrow=c(1,3))
pc.data <- as.data.frame(pc.data)
pc.data$t <- label
pc$p <- as.character(result$cluster)
colnames(pc) <- c("PC1","PC2","PC3","True Cluster","scCAN Cluster" )
pal <- hue_pal()(length(result$cluster)))
colors <- pal[as.numeric(tsne$`scCAN Cluster`)]
scatterplot3d(pc.data[,1:3], pch = 16,
grid=TRUE, box=TRUE,color = colors)
tsne <- as.data.frame(tsne.data)
tsne$t <- label
tsne$p <- as.character(result$cluster)
colnames(tsne) <- c("t-SNE1","t-SNE2","t-SNE3","True Cluster","scCAN Cluster" )
pal <- hue_pal()(length(unique(result$cluster)))
colors <- pal[as.numeric(tsne$`scCAN Cluster`)]
scatterplot3d(tsne[,1:3], pch = 16,
grid=TRUE, box=TRUE,color = colors)
umap <- as.data.frame(umap.data)
umap$t <- label
umap$p <- as.character(result$cluster)
colnames(umap) <- c("UMAP1","UMAP2","UMAP3","True Cluster","scCAN Cluster" )
scatterplot3d(umap[,1:3], pch = 16,
grid=TRUE, box=TRUE,color = colors)
mtext(name,side=3,line=-1.5,outer=TRUE,cex = 1.4)
3D Visualization of latent data with PCA, t-SNE and UMAP for Pollen dataset.
Example code for rare cell type detection
The sampling process is necessary to reduce both time and space complexity, but it can alter the capability of detecting rare cell types. By selecting 5,000 cells from a large dataset, we might end up with insufficient number of rare cells, and therefore reduce the chance of detecting them.
We have developed two strategies to enhance the method’s capability of detecting rare cell types. First, we now allow users to change the parameter samp.size so that they can increase the sample size, thus boosting the method’s capability in detecting rare cell types. Second, we provide an instruction to perform multi-state clustering, i.e., further splitting the clustering results. When a cell type has too few cells, these cells are often mistakenly grouped with other cell types. By further splitting each clusters, we are able to detect rare cell types that would not be possible by performing one-stage clustering.
To demonstrate the efficiency of both solutions, we have tested them on the Zilionis dataset [7]. The Zilionis dataset has 34,558 cells and 9 cell types. Among the 9 cell types, the tRBC cell type has only 108 cells (0.3%). A sub-sample of 5,000 cells is expected to have approximately 19 tRBC cells, which might be insufficient for many clustering method to detect them. In is indeed the case for this dataset. When we use the default value of \(samp.size=5,000\), the tRBC cells are groupped with tPlasma cells (panel B in the figure below).
To demonstrate the efficiency of the first strategy, we set \(samp.size=10,000\). The method is able to seperate tRBC cells (cluster 2 in panel C of the figure below). The method is expected to perform even better if we further increase the sample size.
To demonstrate the efficiency of the second strategy, we performed a two-stage clustering using the the default setting of \(samp.size=5,000\). In stage one, we partitioned the data using scCAN and obtained the clustering results. In stage two, we further partitioned each cluster obtained from stage one using the same method scCAN. Again, as shown in the figure below, the method is able to separate tRBC cells (cluster 2_2 in panel D of the figure below)
The code to detect rare cell types from Zilionis dataset using two proposed approaches are shown below:
suppressPackageStartupMessages({
library(SummarizedExperiment)
})
path <- "https://bioinformatics.cse.unr.edu/software/scCAN/data/zilionis.rds"
SCE <- readRDS(url(path,"rb"))
# Get expression matrix
data <- t(SummarizedExperiment::assay(SCE))
# Get cell annotation
label <- SummarizedExperiment::colData(SCE)
# Detect rare cell types by increasing sample size and reducing number of neighboring cells
res <- scCAN(data, samp.size = 10000, n.neighbors = 10)
# Detect rare cell types by performing two-stage clustering analysis
# Stage 1: Perform clustering on the whole data using scCAN with default settings
res <- scCAN(data)
cluster <- res$cluster
# Stage 2: Perform clustering on each cluster obtain from stage 1
cluster2 <- rep(NA, length(cluster))
for (cl in unique(cluster)){
idx <- which(cl == cluster)
tmp <- scCAN(data[idx, ])
cluster2[idx] <- paste0(cl, "_", tmp$cluster)
}
res$cluster2 <- cluster2We can use the same approach presented above to to visualize the transcriptome landscape with true cell types and clusters obtained from scCAN.
Rare cell type detection result measured by F1 score using original scCAN and two rare cell types detection strategies.
Example code for in-depth analysis using scRNA-seq data and prior biological knowledge.
Download data
In this section, we investigate the scaling ability of scCAN for clustering big data (>1 million cells). We will use neural dataset obtained from multiple brain regions of two embryonic mice published by 10X Genomics [9]. The dataset was stored in DelayedArray format that contains 1,300,774 cells and 27,998 genes. The DelayedArray stores data in physical hard disk and the data can be read to memory by small blocks. We will use TENxBrainData R package available from Bioconductor to efficiently load the data. Before running the following code, users need to make sure that they have sufficient memory (200+ GB of RAM) and data storage (10GB).
# Loading
# if (!require("BiocManager", quietly = TRUE))
# install.packages("BiocManager")
# BiocManager::install("DelayedArray")
# BiocManager::install("DelayedMatrixStats")
# BiocManager::install("TENxBrainData")
# BiocManager::install("HDF5Array")
suppressPackageStartupMessages({
library(TENxBrainData)
library(Matrix)
library(SummarizedExperiment)
library(DelayedArray)
library(DelayedMatrixStats)
library(HDF5Array)
})
tenx <- TENxBrainData()
counts <- assay(tenx)
data <- t(counts)
gene.names <- rowData(tenx)
cell.names <- colData(tenx)
rownames(data) <- cell.names$Barcode
colnames(data) <- gene.names$EnsemblData pre-processing and clustering
The raw data is extremely sparse with excessive number of zero values. We perform one pre-processing step to remove genes that have low expression values. We filter the genes that have total UMI counts across all cells less than 1% of the maximum total UMI counts of all genes.The code for pre-processing is presented as follows:
# Set number of processing unit to 20
setAutoBPPARAM(BiocParallel::MulticoreParam(workers=4))
# Calculate total count for each gene
total.count <- DelayedArray::colSums(data)
# Filter gene that has total count less than 1 percent of maximum total count.
setAutoBPPARAM(BiocParallel::MulticoreParam(workers=4))
idx <- which(total.count <= (max(total.count*1/100)))
data <- data[,-idx]
genes <- as.vector(gene.names$Ensembl)[-idx]
colnames(data) <- genesTo cluster the obtained data, we need to convert it to sparse matrix using the code below:
# Devide data into different block
trunks<- seq(from = 1, to = nrow(data))
seg <- split(trunks, ceiling(seq_along(trunks)/50000))
# Convert each block to sparse dgCMatrix and combine them into a single one
mat = NULL
for(i in 1:length(seg)){
tmp <- as(data[seg[[i]],], "dgCMatrix")
mat = rbind(mat,tmp)
}
rownames(mat) <- cell.names$Barcode
colnames(mat) <- genes
# Set the number of clusters k = 10:50 (k = 2:15 by default) for scCAN to maximize the number clusters that can be explored.
start <- Sys.time()
res <- scCAN(mat,sparse = T,r.seed = 1,k = 10:50)
running_time <- difftime(Sys.time(),start,units = "mins")
cluster <- res$cluster
message("Running time = ", running_time)
message("Number of cluster = ", unique(cluster))## Running time = 51.37449 mins
## Number of cluster = 19scCAN was able to cluster 1.3 million cells within 51 minutes with 19 discovered clusters.
Visualization and clustering results of 1.3 million cells sequenced from brain tissue of E18 mice using scCAN’s clusters.
Clustering result validation using biological analysis
Originally, the 1.3M dataset does not have true cell types. The cluster results were obtained using k-means clustering algorithms with number of clusters k vary from 2 to 20. Therefore, we will validate the sccAN’s clustering results using biologically validated cell types with known marker genes. We retrieve markers genes of 31 neural cell types for mouse brain tissue in PanglaoDB database [10]. We filter the cell types that have a number of markers lower than 15. As the results, we obtain a list of markers genes for 24 cell types. We consider these biologically validated cell types as a ground truth to validate scCAN result.
We follow the approach proposed by Xie et al. [11] to perform validation of scCAN result. For each cluster obtained from scCAN, we divide cells in each gene into two parts: (i) cells belong aforementioned cluster, and (ii) cells belong to other clusters. Next, we perform two-sided Wilcoxon test using two sets of cells with corraction to assess the differential expression. We also calculate fold-change of average expression of two sets to measure ratio of the difference. Output from this step is a list Wilcoxon p-value and fold-change values of all genes for each cluster identified from scCAN. We repeat this project for every gene in all clusters obtained from scCAN. Here, we set the threshold cut-off of the Wilcoxon test (\(p\leq0.001\)) for a strict detection of a small number of marker genes and 1.5 for fold-change. The codes to process markers genes and conduct differential analysis are shown below:
# Download the cell type with markers gene
url <- "https://panglaodb.se/markers/PanglaoDB_markers_27_Mar_2020.tsv.gz"
tmp <- tempfile()
download.file(url,tmp)
file = gzfile(tmp)
markers <- read.table(file, sep="\t", header = T)
# Select only brain tissue
markers = markers[which(markers$organ=="Brain"),]
# Remove marker that belongs to human
markers = markers[c(which(markers$species=="Mm Hs"),which(markers$species=="Mm")),]
# Select cell types that have more than 15 markers
breaks = c(15)
for(br in breaks){
tmp <- as.data.frame(table(markers$cell.type))
if(is.null(br)){
select_mrks <- tmp
}else{
select_mrks <- tmp[which(tmp$Freq>=br),]
}
cts <- unique(select_mrks$Var1)
panglao_markers <- list()
# ann <- NULL
for( ct in cts){
idx <- which(markers$cell.type==ct)
tmp <- markers[idx,]
panglao_markers[[ct]]<-tmp$official.gene.symbol
}
saveRDS(panglao_markers,paste0("Panglao_mrks_",br,".rds"))
}
# Load the expression matrix of markers genes collected from PanglaoDB database
path <- "https://bioinformatics.cse.unr.edu/software/scCAN/data/1.3M_markers.rds"
gene_expression_data <- readRDS(url(path,"rb"))
# Load saved scCAN clustering result
path <-"https://bioinformatics.cse.unr.edu/software/scCAN/result/1.3M_scCAN.rds"
res <- readRDS(url(path,"rb"))
cluster <- res$cluster
# Evaluate wilcox rank sum test and log fold change values
# Perform pair wise Wilcoxon Rank Sum and log fold change for each cluster
# This process could take hours for big data
eval_gene_markers <- scCAN::get_cluster_markers(gene_expression_data,
cluster,
threads = 16)
gene_names <- rownames(gene_expression_data)
# Select marker the test with p <= 0.001 and log fold change > 1.5,
# output is the list of markers that are strongly differentiated in the clusters
cluster_markers_list <- scCAN::curate_markers(eval_gene_markers, gene_names,
wilcox_threshold=0.001,
logfc_threshold=1.5)Next, we calculate the probability of a clusters belong to a reference cell type. Output from this step is a confusion matrix where rows represent scCAN’s clusters and columns represent reference cell types.
library(matrixStats)
# Load pre-saved the list of markers that are strongly expressed in a specific cluster
path <-"https://bioinformatics.cse.unr.edu/software/scCAN/result/cluster_markers.rds"
cluster_markers_list <- readRDS(url(path,"rb"))
# Load PanglaoDB markers
path <-"https://bioinformatics.cse.unr.edu/software/scCAN/result/panglao_markers.rds"
panglao_markers<- readRDS(url(path,"rb"))
# Caculate pair wise cell type/cluster probabilty using jaccard index
celltype_prob <- scCAN::calculate_celltype_prob(cluster_markers_list,
panglao_markers,
type = "jacc")
colnames(celltype_prob) <- names(panglao_markers)
celltype_id <- matrixStats::rowMaxs(celltype_prob)
celltype_prob[1:10,1:10] Anterior pituitary gland cells Astrocytes Bergmann glia Cajal-Retzius cells Choroid plexus cells
[1,] 0.01190476 0.00000000 0.00000000 0.04205607 0.02564103
[2,] 0.03050847 0.12962963 0.08278146 0.02052786 0.02473498
[3,] 0.00000000 0.00000000 0.00000000 0.00000000 0.00000000
[4,] 0.04216867 0.08205128 0.05780347 0.03773585 0.01298701
[5,] 0.02631579 0.01398601 0.01652893 0.00625000 0.00000000
[6,] 0.01754386 0.00000000 0.00000000 0.02912621 0.02222222
[7,] 0.02756892 0.10981308 0.07389163 0.04269663 0.02842377
[8,] 0.02546296 0.02386117 0.01594533 0.08577406 0.01428571
[9,] 0.03398058 0.02494331 0.01193317 0.15283843 0.02000000
[10,] 0.03030303 0.03608247 0.03488372 0.13270142 0.02614379
Dopaminergic neurons GABAergic neurons Immature neurons Interneurons Meningeal cells
[1,] 0.03246753 0.006711409 0.04651163 0.12146893 0.000000000
[2,] 0.01423488 0.014492754 0.02006689 0.09147609 0.017921147
[3,] 0.00000000 0.000000000 0.00000000 0.00625000 0.000000000
[4,] 0.00000000 0.013605442 0.06470588 0.06534091 0.013333333
[5,] 0.02000000 0.063157895 0.02542373 0.08333333 0.010204082
[6,] 0.00000000 0.026315789 0.06557377 0.02880658 0.000000000
[7,] 0.01038961 0.010526316 0.02729529 0.10769231 0.020887728
[8,] 0.01913876 0.009685230 0.03211009 0.16990291 0.007211538
[9,] 0.01758794 0.005089059 0.01923077 0.17224080 0.012626263
[10,] 0.02649007 0.013698630 0.02366864 0.12535613 0.013422819We can view the cluster that is strongly correlated with the reference cell type using heatmap plot. Color scale is reported by the value of celltype_prob for each cluster-label combination.
library(pheatmap)
pheatmap(celltype_prob)
We select the cluster that has highest probability value with reference cell type.
path <-"https://bioinformatics.cse.unr.edu/software/scCAN/result/1.3M_scCAN.rds"
res <- readRDS(url(path,"rb"))
cluster <- res$cluster
# Map the cluster that has highest Jaccard Index value with reference cell type
celltype_idx <- apply(celltype_prob,1,function(x){which.max(x)} )
cell.name <- colnames(celltype_prob)[celltype_idx]
# Assign reference cell type to all clusters discover by scCAN
mapping.table <- data.frame(cluster = seq(1:19), cell_type = cell.name)
mapped.ct <- mapping.table$cell_type[mapping.table$cluster[cluster]]
mapped.ct[1:10][1] "Interneurons" "Pyramidal cells" "Astrocytes" "Interneurons" "Neurons" "Neurons"
[7] "Pyramidal cells" "Neurons" "Pyramidal cells" "Pyramidal cells"