new_clustering.R

#Libraries
library( Matrix )
library( tidyverse )
library( rhdf5 )
library( HDF5Array )

readTENxH5File <- function( filename )
  sparseMatrix(
    p = h5read( filename, "matrix/indptr" ),
    i = h5read( filename, "matrix/indices" ) + 1,
    x = as.numeric( h5read( filename, "matrix/data" ) ),
    dims = h5read( filename, "matrix/shape" ),
    dimnames = list(
      as.vector( h5read( filename, "matrix/gene_names" ) ), 
      as.vector( h5read( filename, "matrix/barcodes" ) )
    )
  )

raw <- readTENxH5File( "~/sds/sd17l002/u/anders/tmp/ASD.h5" )
rawt <- t(raw)

cs <- colSums(raw)

meta <- read_tsv("~/sds/sd17l002/u/anders/tmp/ASD2/meta.txt" )

meta %>% 
select( sample : `RNA Integrity Number` ) %>%
unique ->
  sampleTable

meta %>% 
select( -( individual : `RNA Integrity Number` ) )  ->
  cellTable

stopifnot( all( str_sub( cellTable$cell, 1, 14 ) == colnames(raw) ) )
stopifnot( all( str_sub( cellTable$cell, 1, 14 ) == rownames(rawt) ) )

colVars_spm <- function( spm ) {
  stopifnot( is( spm, "dgCMatrix" ) )
  ans <- sapply( seq.int(spm@Dim[2]), function(j) {
    mean <- sum( spm@x[ (spm@p[j]+1):spm@p[j+1] ] ) / spm@Dim[1]
    sum( ( spm@x[ (spm@p[j]+1):spm@p[j+1] ] - mean )^2 ) +
      mean^2 * ( spm@Dim[1] - ( spm@p[j+1] - spm@p[j] ) ) } ) / ( spm@Dim[1] - 1 )
  names(ans) <- spm@Dimnames[[2]]
  ans
}

gene_means <- colMeans( rawt/cs )
gene_vars <- colVars_spm( rawt/cs )
xi <- mean( 1/cs )

plot( gene_means, gene_vars / gene_means, pch=".", log="xy" )
abline( h = 1.5*xi )

infrmgenes <- gene_vars / gene_means > 1.5*xi & gene_means > 3e-8
rawi <- raw[ infrmgenes, ]

pca <- irlba::prcomp_irlba( sqrt(t(rawi)/cs), 30 )
ump <- uwot::umap( pca$x, metric="cosine", ret_nn=TRUE, n_threads=40, min_dist=0.03, verbose=TRUE )

nn <- ump$nn$cosine$idx[,-1]
ump <- ump$embedding

plot( ump, pch=".", col=adjustcolor( as.integer(as.factor(cellTable$cluster)), .1 ), asp=1 )
text( tapply( ump[,1], cellTable$cluster, median ), tapply( ump[,2], cellTable$cluster, median ), 
      levels(as.factor(cellTable$cluster)) )


# Get Louvain clusters

library( igraph )
g <- graph_from_edgelist( as.matrix( map_dfr( 1:10, function(i) data.frame( from=1:nrow(nn), to=nn[,i] ) ) ) )
louv <- cluster_louvain( as.undirected( g, "collapse" ) )

ncl <- max(membership(louv))

plot( ump, pch=".", col = rainbow( ncl, alpha=.1, v=.7 )[ membership(louv) ], asp=1 )
text( tapply( ump[,1], membership(louv), median ), tapply( ump[,2], membership(louv), median ), 1:ncl )


# Border cells: How many of the k nearest neighbors of a cell are not in the same cluster as the cell?
foreignNN <- rowSums( matrix( membership(louv)[ nn ], ncol=ncol(nn) ) != membership(louv) )
table( foreignNN )

# These cells have no foreign neighbors
interior <- foreignNN == 0

plot( ump, pch=".", col = adjustcolor( ifelse( interior, "black", "red" ), .1 ), asp=1 )

# For which cells is the first nearest neighbor another cluster and which is it?
tibble( cell_cluster = membership(louv), nn1_cluster = membership(louv)[nn[,1]] ) %>%
  group_by_all %>% tally %>% spread( nn1_cluster, n, fill=0 ) %>% 
  column_to_rownames("cell_cluster") -> cluster_nbgh

# Show graph of clusters with at least x% of cells having a neighbor in the other cluster
plot( graph_from_adjacency_matrix( cluster_nbgh / sizes(louv) > .01 ) )

comp <- components( graph_from_adjacency_matrix( cluster_nbgh / sizes(louv) > .02 ) )
gcms <- comp$membership[ membership(louv)  ] # grand cluster membership
foreignNN <- rowSums( matrix( gcms[ nn ], ncol=ncol(nn) ) !=gcms )

plot( ump, pch=".", col = ifelse( foreignNN>0, "#00000018", rainbow( ncl, alpha=.1, v=.7 )[ gcms ] ), asp=1 ) 
text( tapply( ump[,1], gcms, median ), tapply( ump[,2], gcms, median ), 1:20 )

#Calculation of pseudobulks for one cluster at a time
#Add new clustering to cellTable as new column

cellTable <- cellTable %>%
  mutate( newcluster = sprintf( "MC%02d", gcms ) )

#Relate new clusters to old clusters
tibble( clusterkey= c( "Oligodendrocytes", "L4 / L5/6-CC", "OPC", 
                     "Astrocytes", "Neu NRGN", "L2/3 / Neumat", "Astrocytes", 
                     "IN-PV", "IN-SST", "IN-VIP / IN-SV2C", "L5/6", "L5/6",
                     "Endothelial", "IN-PV", "Microglia", "L5/6-CC", "L4",
                     "AST-PP", "IN-SV2C" ) , newcluster = sort( unique( cellTable$newcluster ) ) ) %>%
  right_join( cellTable, by="newcluster" )


counts <- TENxMatrix( "/home/anders/pub/ASD.h5", "matrix" )
do_DESeq_on_newcluster <- function( cluster ){
  pseudobulk <- sapply( sampleTable$sample, function(s)
    rowSums( counts[ , meta$sample == s & cellTable$newcluster==cluster, drop=FALSE ] ) )
  dds <- DESeqDataSetFromMatrix( pseudobulk, sampleTable, ~ sex+region+age+diagnosis)
  dds <- dds[ rowSums(counts(dds)) >= 10, colSums(counts(dds)) > 0 ]
  dds <- DESeq( dds )
  dds
}

plan( multiprocess, workers=20 )
dds_nc <- future_map( sort( setdiff( unique( cellTable$newcluster ), "MC18" ) ), do_DESeq_on_newcluster )
names( dds_nc ) <- sort( setdiff( unique( cellTable$newcluster ), "MC18" ) )
      
genenames <- tibble( ensg = h5read( "ASD.h5", "matrix/genes" ), 
        name = h5read( "ASD.h5", "matrix/gene_names" ) )
res.paper <- read.delim( "~/sds/sd17l002/u/isabel/S4.csv", sep=";", dec = "," ) %>%
  select( cluster = Cell.type, gene = gene.ID, name=Gene.name, q.value )

res.tibble <- map_dfr( dds, as_tibble, rownames="gene", .id="cluster" ) %>%
  left_join( genenames, by=c( "gene" = "ensg" ) ) #%>%
  #filter( .$padj < .1 )

res.nc <- map_dfr( dds_nc, function(x) x %>% results %>% as_tibble( rownames="gene" ), .id="newcluster" ) %>%
  left_join( genenames, by=c( "gene" = "ensg" ) )
 

#Plot that relates significance from DESeq output to MAST output
res.tibble %>%
  left_join( res.paper )%>%
  mutate(signif.DESeq = !is.na( padj ) & padj < .05, 
         signif.MAST = !is.na( q.value ) & q.value <.05 ) %>% 
  ggplot+
      geom_point( aes(signif.MAST, signif.DESeq, col=signif.DESeq ), size=.1, position= "jitter" )+
      facet_wrap( ~cluster )

#Plot that compares DESeq on manual clusters and paper clusters
#Currently this plots all genes, which takes long and is not very smart
res.nc %>%
  mutate( padjnc = .$padj,
            padj=NULL ) %>%
  left_join( res.tibble, by="gene" )%>%
  mutate( sign.nc = padjnc < .1, sign.pc = padj <.1 ) %>% 
ggplot+
  geom_point( aes( sign.nc, sign.pc, col=sign.nc ), size=.1, position= "jitter" )+
  facet_wrap( ~newcluster )

#look at diagnostic marker genes for different EN layers
  ggplot()+
    geom_point( aes( ump[, 1], ump[,2], col= raw[ "TLE4", ]/cs ), size=.01 )+
    scale_color_gradientn( colours = rev( rje::cubeHelix( 100 ) ), trans="sqrt" )+
    geom_text( aes( tapply( ump[,1], cellTable$cluster, median ), 
                    tapply( ump[,2], cellTable$cluster, median ), 
              label=levels( factor( cellTable$cluster ) ) ), data=NULL )
  
  
#Temporary object with raw and normalised counts
k <-   tibble( NONO = raw[ "NONO", ], cell = meta$cell, fracMT = raw[ "MTND2P28", ]/cs , 
              BAG3 = raw[ "BAG3", ], GJA1 = raw[ "GJA1", ], fracNONO = raw[ "NONO", ]/cs, 
              fracGJA1 = raw[ "GJA1", ]/cs, TCF25 = raw[ "TCF25", ], fracTCF25 = raw[ "TCF25", ], 
              TTF2 = raw["TTF2", ], fracTTF2 = raw["TTF2", ]/cs,
              newcluster = cellTable$newcluster, clusterkey = cellTable$clusterkey )%>%
         left_join( meta, by="cell" )


ggplot(k, aes( k$fracMT, k$diagnosis, col=k$diagnosis ) )+
  geom_point( position = "jitter", size=.1 )+
  facet_wrap( ~k$sample )
  
ggplot(k, aes( sqrt( k$fracNONO ), k$sample, col=k$diagnosis ) )+
    geom_point( position = "jitter", size=.1 )+
    facet_wrap( ~k$newcluster )
  
k  %>% 
  group_by( sample, diagnosis, newcluster, cs ) %>% 
  filter(newcluster=="MC08", log10(cs) > 3.5 ) %>%
  summarise( y=mean( sqrt( NONO ) > 1  ) ) %>% 
  ggplot() + 
  geom_point( aes( x=diagnosis, y=y ), position="jitter", size=.2 )+
  facet_wrap(~newcluster)
  
  
k %>% 
  #add_column(cs) %>% 
  dplyr::filter(newcluster=="MC08") %>% 
  mutate( y=NONO+runif( n(), 0, .5) )%>% 
  ggplot() + 
  geom_point(aes(x=log10(cs),y=y,col=diagnosis), size=.2) + 
  scale_y_continuous(trans="sqrt")

  
  k %>% add_column( cs ) %>% 
    #filter( newcluster=="MC06" ) %>%
    {ggplot(.)+
    geom_density( aes( sqrt( .$fracTTF2 ), col=.$diagnosis ) )+
    geom_vline( xintercept = .$location ) +
    facet_wrap(~.$newcluster)}
  
#Implementing locmodes to see if that can produce sensible cut-point
  k <- k %>%
    group_by(newcluster) %>%
    summarise(location = multimode::locmodes( sqrt( fracTTF2 ), 2 )$location[2]) %>%
    right_join(k)
 
#
  l <- k%>%
    dplyr::filter(newcluster=="MC08") %>%
    mutate(NONO=NONO+runif(n(),0,.4))
 ggplot()+
   geom_point( aes( log10( cs ), sqrt( NONO ) ), size=.2, data=dplyr::select(l, NONO, cs), col="grey" )+
    geom_point(aes(log10( cs ), sqrt( NONO ), col=diagnosis ), size=.2, data=l)+
   facet_wrap(~sample)
 
 
 k %>% 
   dplyr::filter( newcluster=="MC08" ) %>% 
   group_by( diagnosis, sample ) %>% 
   summarise( m = mean( fracNONO ) ) %>% 
   ggplot + 
   geom_point(aes(x=sample,y=m,col=diagnosis))+
   theme(axis.text.x = element_text(angle = 90))
 

  
#Look at GO
 library(clusterProfiler)
 cluster <- "MC18"
 h <- res.nc %>% dplyr::filter(newcluster== cluster & baseMean > 5) %>% dplyr::pull( log2FoldChange )
 k <- res.nc %>% dplyr::filter(newcluster== cluster & baseMean > 5) %>% dplyr::pull( gene )
 #Instead of filtering baseMean, excluding genes with padj of NA also works

 ## feature 1: numeric vector
 geneList <- h
 
 ## feature 2: named vector
 names(geneList) <- as.character(k)
 gene <- res.nc %>% dplyr::filter(newcluster == cluster & padj<.1 )  %>% pull(gene)
 
 ## feature 3: decreasing order
 geneList <- sort(geneList, decreasing = TRUE)

 ego <- enrichGO(gene         = gene,
                  universe      = names(geneList),
                  OrgDb         = org.Hs.eg.db,
                  keyType       = 'ENSEMBL',
                  ont           = "BP",
                  pAdjustMethod = "BH",
                  pvalueCutoff  = 0.01,
                  qvalueCutoff  = 0.05)
ego

#Transcribe Ensemble to go id or likewise
mart <- useMart("ensembl", "hsapiens_gene_ensembl")
a <- getBM( c("ensembl_gene_id","go_id"), mart=mart )

#Look at overlap to GWAS Studies
SFARI <- read.csv("~/sds/sd17l002/u/isabel/SFARI-Gene_genes_08-20-2019release_08-23-2019export.csv")

s <- res.nc %>%
  dplyr::filter(padj < .1 & !is.na(padj) ) %>%
  dplyr::pull(name) %>%
intersect(SFARI$gene.symbol)
 

res.nc%>%
  dplyr::filter(res.nc$name %in% s & !is.na(padj) & padj <.1 ) %>%
  View()
SFARI %>%
  as_tibble() %>%
  dplyr::filter(gene.symbol %in% s) %>%
  View()