In vondoRishi/scpackages: scpackages has many single-cell R and python packages to use them easily.

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(scpackages)

Load Seurat Object

First we will load a scaled and normalized seurat object from disk.

scaled_cluster <- readRDS("../scaled_cluster.13_February_2020.rds")

Gene ~ PC

Which genes are loaded in which principal component?

tmp_hvf <-HVFInfo(scaled_cluster)
tmp_hvf$Symbol <- rownames(tmp_hvf)
# write.csv(tmp_hvf, file = "HVF_info.csv",row.names = FALSE)

## gene PCA Explore ##
scaled_cluster@misc$loading <- scpackages::pca_gene_loading(scaled_cluster)
scaled_cluster@misc$loading <-
  scaled_cluster@misc$loading %>% left_join(tmp_hvf, by = "Symbol")

library(ggrepel)

ggplot(scaled_cluster@misc$loading, 
       aes(x=factor(pca_fix),y=variance.standardized, label = Symbol)) +
  geom_point(color = ifelse(scaled_cluster@misc$loading$variance.standardized > 4 
                            , "red", "black")) +
  geom_text_repel(
    data = subset(scaled_cluster@misc$loading, variance.standardized > 6 )
  )+ggtitle(label = "Visualize the gene loading in each PCA ")

Parameters

There are two parameters pca and resolution along with number of HVGs which controls number of cell types (cluster)

Elbow plot

scaled_cluster@misc$elbowPlot+ggtitle("Where is the elbow")
## ??Where is the elbow does it matters ??

What are the optimal number of PC and resolution value?

featureRange <- seq(200,1000, by=400)
featureRange <- c(50,100)
pcRange <- seq(15,25,by = 5)
resolutionRange <- seq(0.4,1.2,by = 0.4)

scaled_cluster <- scpackages::gridFindClusters( Object = scaled_cluster, 
                                                pcRange, resolutionRange, featureRange,
                                                identPrefix = "solution", 
                                                verbose =FALSE)

ggplot(scaled_cluster@misc$data_NClust,
       aes(x = resolution, y = Cell_types, color=factor(feature) )) + geom_line() + facet_wrap(. ~ pc)

Keeping "PCA" constant, higher resolution provides higher number of cell types

Increasing number of PCA decreases max number of cell types

With 17 PCA , resolution (0.4, 0.6,0.8) produces 6, 8 and 9 clusters

Are they belong to same solution?

library(easyalluvial)
alluvial_wide(dplyr::select(
  scaled_cluster[[]],
  solution_17_0.4,
  solution_17_0.6,
  solution_17_0.8
)) + theme_bw()

Cluster the similar solutions?

scaled_cluster <-  scpackages::findSimilarClusterSolution(scaled_cluster,
                                         identPrefix = "solution",
                                         similarityCut = 0.8)
plot(scaled_cluster@misc$cluster_G1, vertex.color = 
       rainbow(10, alpha = 0.8)[scaled_cluster@misc$clusterlouvain$membership])

Look for same shape( number of cluster) with different color(solution)

ggplot(  scaled_cluster@misc$meta_cluster,  aes(
                y = resolution,
                x = pca,
                # shape = factor(Cell_types),
                label = Cell_types,#membership,
                color = factor(membership)
              )
)  + geom_text()
#+  scale_color_manual(values =  rainbow(8)) #+  theme_dark()

Finding different solutions with same number of cell types.

cluster_representatives <- as_tibble( scaled_cluster@misc$meta_cluster %>% 
                                        group_by(membership,Cell_types) %>% 
                                        sample_n(1)) %>% arrange(pca,resolution)


x <- as.list(cluster_representatives %>% 
               filter( Cell_types==9 ) %>% 
               select(cluster_id))
alluvial_wide(scaled_cluster[[]][, as.character(x$cluster_id)]) + 
  theme(axis.text.x =  element_text(angle = 90)) +
  theme_bw() + 
  ggtitle(paste(    length(x$cluster_id),
    " different solutions with equal number of clusters"
  ))

Here, we can see there are 4 different clustering solutions although each of them give exact 9 cell types. For further discussion, solution_15_0.8 is marked as reference set. Now the cell_types 3, 5, 6, 7 and 8 are always intact. Although, other cell_types are not constant in different solutions, but after subdividing into sub-group they remain mostly together.

vondoRishi/scpackages documentation built on Feb. 27, 2020, 1:52 a.m.