Homepage
The formatted source code for this file is here.
And a raw version here.
Previous work by Youngser Park can be found here.

1 Data

Here we read in the data and select a random half of it for exploration.

featFull <- fread("../data/synapsinR_7thA.tif.Pivots.txt.2011Features.txt",showProgress=FALSE)

### Setting a seed and creating an index vector
### to select half of the data
set.seed(2^10)
half1 <- sample(dim(featFull)[1],dim(featFull)[1]/2)
half2 <- setdiff(1:dim(featFull)[1],half1)

feat <- featFull[half1,]
dim(feat)
# [1] 559649    144
## Setting the channel names
channel <- c('Synap_1','Synap_2','VGlut1_t1','VGlut1_t2','VGlut2','Vglut3',
              'psd','glur2','nmdar1','nr2b','gad','VGAT',
              'PV','Gephyr','GABAR1','GABABR','CR1','5HT1A',
              'NOS','TH','VACht','Synapo','tubuli','DAPI')

## Setting the channel types
channel.type <- c('ex.pre','ex.pre','ex.pre','ex.pre','ex.pre','in.pre.small',
                  'ex.post','ex.post','ex.post','ex.post','in.pre','in.pre',
                  'in.pre','in.post','in.post','in.post','in.pre.small','other',
                  'ex.post','other','other','ex.post','none','none')

channel.type2 <- c('ex.pre','ex.pre','ex.pre','ex.pre','ex.pre','other',
                   'ex.post','ex.post','ex.post','ex.post','in.pre','in.pre',
                   'in.pre','in.post','in.post','in.post','other','other',
                   'ex.post','other','other','ex.post','other','other')
nchannel <- length(channel)
nfeat <- ncol(feat) / nchannel

## Createing factor variables for channel and channel type sorted properly
ffchannel <- (factor(channel.type,
    levels= c("ex.pre","ex.post","in.pre","in.post","in.pre.small","other","none")
    ))
fchannel <- as.numeric(factor(channel.type,
    levels= c("ex.pre","ex.post","in.pre","in.post","in.pre.small","other","none")
    ))
ford <- order(fchannel)


## Setting up colors for channel types
Syncol <-  c("#197300","#5ed155","#660000","#cc0000","#ff9933","#0000cd","#ffd700")
Syncol3 <- c("#197300","#197300","#cc0000","#cc0000","#0000cd","#0000cd","#0000cd")
ccol <- Syncol[fchannel]
ccol3 <- Syncol3[fchannel]

exType <- factor(c(rep("ex",11),rep("in",6),rep("other",7)),ordered=TRUE)
exCol<-exType;levels(exCol) <- c("#197300","#990000","#0000cd");
exCol <- as.character(exCol)

fname <- as.vector(sapply(channel,function(x) paste0(x,paste0("F",0:5))))
names(feat) <- fname
fcol <- rep(ccol, each=6)
mycol <- colorpanel(100, "purple", "black", "green")
mycol2 <- matlab.like(nchannel)
mycol3 <- colorpanel(100, "blue","white","red")

1.1 Data transformations

f <- lapply(1:6,function(x){seq(x,ncol(feat),by=nfeat)})
featF <- lapply(f,function(x){subset(feat,select=x)})

featF0 <- featF[[1]]
f01e3 <- lapply(featF,function(x){
                  1e3*data.table(apply(X=x,2,function(y){((y-min(y))/(max(y)-min(y)))}))
                   })

fs <- f01e3[[1]]

### Taking log_10 on data + 1.
log1f <- log10(featF0 + 1)
slog1f <- data.table(scale(log1f, center=TRUE,scale=TRUE))

We now have the following data sets:

  • featF0: The feature vector looking only at the integrated brightness features.
  • fs: The feature vector scaled between \([0,1000]\).
  • logf1: The feature vector, plus one, then \(log_{10}\) is applied.
  • slog1f: The feature vector, plus one, \(log_{10}\), then scaled by subtracting the mean and dividing by the sample standard deviation.

2 Marker Exploration

2.1 Correlation Matrix of markers

#corrf <- lapply(lapply(f01e3, cor), function(x) x[ford, ford])
corrf <- lapply(lapply(featF, cor), function(x) x[ford, ford])

titles <- paste('Correlation Matrix of', paste0("F", 0:5))
par(mfrow = c(3,2))
for(i in 1:length(corrf)) {
  corrplot(corrf[[i]],method="color",tl.col=ccol[ford], 
           tl.cex=0.8, mar=c(1,0,1,1.5))
  title(titles[i])
}
Figure 1: Correlation on untransformed F0 data, reordered by synapse type.

bford <- order(rep(fchannel,each=6))
nord <- Reduce('c', f)
cr <- rep(ccol, each=6)
corrfB <- cor(feat)[bford,bford]
corrplot(corrfB,method="color",tl.col=cr[bford],tl.cex=0.75)
Figure 2: Full Correlation on untransformed F0-F5 data, reordered by synapse type.

2.2 Distance Correlation T-tests

computeDcor <- function(x) {
  set.seed(317)
  sam1 <- sample(dim(x)[1], min(1e3,dim(x)[1]))
  tmp <- as.data.frame((x[sam1,]))
  
  combcols <- t(combn(24,2))
  
  dc <- foreach(i = 1:dim(combcols)[1]) %dopar% {
         set.seed(331*i)
         dcor.ttest(x=tmp[,combcols[i,1]],y=tmp[,combcols[i,2]])
         }
  
  ms <- matrix(as.numeric(0),24,24)
  mp <- matrix(as.numeric(0),24,24)
  
  for(i in 1:length(dc)){
      ms[combcols[i,1],combcols[i,2]] <- dc[[i]]$statistic
      ms[combcols[i,2],combcols[i,1]] <- dc[[i]]$statistic
      mp[combcols[i,1],combcols[i,2]] <- dc[[i]]$p.val
      mp[combcols[i,2],combcols[i,1]] <- dc[[i]]$p.val
  }
  
  rownames(ms) <- colnames(x)
  rownames(mp) <- colnames(x)
  colnames(ms) <- colnames(x)
  colnames(mp) <- colnames(x)
  
  diag(ms) <- as.numeric(0)
  diag(mp) <- as.numeric(1)
  return(list(ms, mp))
}

mdcor <- lapply(featF, computeDcor)
cl5 <- colorRampPalette(c("white", "blue"))
gr5 <- colorRampPalette(c("darkgreen", "white", "white"))
bl5 <- colorRampPalette(c("blue", "red"))

sTitle <- paste("dcor.ttest statistic", paste0("F", 0:5))
pTitle <- paste("dcor.ttest log10(p-value)", paste0("F", 0:5))


par(mfcol=c(6,2), oma=2*c(1,1,1,1))
for(i in 1:length(mdcor)){
  corrplot(mdcor[[i]][[1]][ford,ford],is.corr=FALSE,method="color",
           tl.col=ccol[ford], tl.cex=0.8, mar=3*c(1,0,1,1.5))
  corrRect(as.numeric(table(fchannel)),col=Syncol,lwd=4)
  title(sTitle[i]) 
}

for(i in 1:length(mdcor)){
  corrplot((log10(mdcor[[i]][[2]][ford,ford]+.Machine$double.eps)),is.corr=FALSE,method="color",tl.col=ccol[ford], 
           tl.cex=0.8, mar=3*c(1,0,1,1.5),
           p.mat=mdcor[[i]][[2]][ford,ford], 
           sig.level=0.01, pch = 'x', pch.cex=0.5)
  corrRect(as.numeric(table(fchannel)),col=Syncol,lwd=4)
  title(pTitle[i])
}
Figure 3: F0-5: Distance correlation t-test statistic and p-value matrices.

The X’s in the above right figure denote a p-value greater than 0.01.

2.3 PCA Scatterplots

We will run PCA on the untransformed correlation matrix so the data can be viewed in 2-dimensions. The colors correspond to synapse type.

pcaL <- lapply(corrf, prcomp, center=TRUE, scale=TRUE)
titlepca <- paste("PCA on ", paste0("cor(F", 0:5, ')'))
for(i in 1:length(pcaL)) { 
  pairs(pcaL[[i]]$x[,1:3], col=ccol3[ford],pch=20, cex=2,
        main=titlepca[i])
}
Figure 4: PCA of untransformed correlation matrix

Figure 4: PCA of untransformed correlation matrix

Figure 4: PCA of untransformed correlation matrix

Figure 4: PCA of untransformed correlation matrix

Figure 4: PCA of untransformed correlation matrix

Figure 4: PCA of untransformed correlation matrix

2.3.1 PCA on cor(F0)

pca <- pcaL[[1]]$x
rgl::plot3d(pca[,1],pca[,2],pca[,3],type='s',col=ccol3[ford], size=1,
            xlab = "PC1", ylab = "PC2", zlab = "PC3")
rgl::rgl.texts(pca[,1],pca[,2],pca[,3],abbreviate(rownames(pca)), col=ccol3[ford], adj=c(0,2))
title3d(main = "PCA on cor(F0)")
subid <- currentSubscene3d()
rglwidget(elementId="rgl-pca1",width=720,height=720)
pcaB <- prcomp(corrfB,center=TRUE, scale=TRUE)
pairs(pcaB$x[,1:3], col=cr[bford],pch=20, cex=2)
Figure 6: PCA of untransformed correlation matrix

plot(pcaB$x[,1:3], col=cr[bford],pch=20, cex=2)
text(pcaB$x[,1:2], labels=rownames(pcaB$x), pos=4, col=cr[bford])
Figure 7: PC2 v PC1 of untransformed correlation matrix

pcaB <- pcaB$x
rgl::plot3d(pcaB[,1],pcaB[,2],pcaB[,3],type='s',col=cr[bford], size=1)
rgl::rgl.texts(pcaB[,1],pcaB[,2],pcaB[,3],abbreviate(rownames(pcaB)), col=cr[bford], adj=c(0,2))
subid <- currentSubscene3d()
rglwidget(elementId="rgl-pcaB",width=720,height=720)

3 LDA

3.1 Scree Plots

First we will take a look at the scree plots for each of the primitives along with their corresponding elbows as given by Zhu and Ghodsi 2006.

Figure 9: Scree plots for F0-F5

3.2 Cumulative Variance plots

par(mfrow= c(2,3))
for(i in 1:length(pcaL)){
  zs <- rep(1, dim(pcaL[[i]]$x)[2])
  zb <- rep(0, dim(pcaL[[i]]$x)[2])
  zs[el2[[i]]] <- 2
  zb[el2[[i]]] <- 2
  plot(100 * cumsum(pcaL[[i]]$sdev) / sum(pcaL[[i]]$sdev), 
       axes = TRUE,
       type='b', pch = 21, 
       bg = zb, col = zs, 
       xlab = 'd', ylab = '% Var',
       main = paste0('% Var Explained for F', i - 1))
  axis(side = 1, at = el2[[i]], padj = 1.25)
}
Figure 10: % Variance Explained for F0-F5

dat <- lapply(pcaL, function(x) data.frame(x$x))
type <- truth <- factor(exType, ordered = FALSE)
tr <- as.numeric(truth)

lda.fit <- list()

lda.fit[[1]] <- lda(type ~ ., data = dat[[1]][,1:el2[[1]][2]])
lda.fit[[2]] <- lda(type ~ ., data = dat[[2]][,1:el2[[2]][2]])
lda.fit[[3]] <- lda(type ~ ., data = dat[[3]][,1:el2[[3]][2]])
lda.fit[[4]] <- lda(type ~ ., data = dat[[4]][,1:el2[[4]][2]])
lda.fit[[5]] <- lda(type ~ ., data = dat[[5]][,1:el2[[5]][2]])
lda.fit[[6]] <- lda(type ~ ., data = dat[[6]][,1:el2[[6]][1]])

lda.pred <- lapply(lda.fit, predict)
Lm <- 
foreach(i = 1:6, .combine=rbind) %:% 
foreach(j = 1:24, .combine=rbind) %do% {
  out <- NULL
  try({
    set.seed(12)
    ldapred <- as.numeric(predict(lda(type ~ ., dat[[i]][1:j]))$class)
    er <- 1/24 * sum(ldapred != as.numeric(truth))
    out <- data.frame(Lhat = er, d = j, feat = as.factor(i - 1))
    }, silent=TRUE)
  out
}

rownames(Lm) <- NULL

Lm <- data.table(Lm)

Lcvm <- 
foreach(i = 1:6, .combine=rbind) %:% 
foreach(j = 1:24, .combine=rbind) %do% {
  out <- NULL
  try({
    set.seed(12)
    ldapred <- as.numeric((lda(type ~ ., dat[[i]][1:j], CV = TRUE))$class)
    er <- 1/24 * sum(ldapred != as.numeric(truth))
    out <- data.frame(Lhat = er, d = j, feat = as.factor(i - 1))
    }, silent=TRUE)
  out
}

rownames(Lcvm) <- NULL

Lcvm <- data.table(Lcvm)


Lm$CV <- "Without CV"
Lcvm$CV <- "With CV"

tmp <- data.table(rbind(Lm, Lcvm))
tmp$CV <- factor(tmp$CV)
tmp$feat <- factor(tmp$feat)

size <- foreach(i = 1:6, .combine=c) %do% {
  size <- rep(1, 23)
  size[el2[[i]]] <- 2
  size
}

tmp$size <- factor(c(size,size))
levels(tmp$size) <- c("non-elbow", "elbow")

3.3 Misclassification rates for LDA w/ and w/o CV.

xvalVr <- 
ggplot(tmp, aes(x = d, y = Lhat, color = feat)) + 
  facet_grid(. ~ CV) +
  geom_line(alpha = 0.5) +
  geom_point(alpha = 0.65, aes(size = size)) + 
  geom_hline(yintercept = 0.64236) +
  ggtitle("FLD (LDA): xval vs. re-substitution")

print(xvalVr)
Figure 11: Lhat vs \(d\) for FLD on F0-F5 with an h-line at chance

3.4 Misclasifications vs \(\hat d\).

L <- 
foreach(i = 1:6) %:% 
foreach(j = 1:24, .combine = cbind) %do% {
  out <- NULL
  try({
    set.seed(12)
    ldapredcv <- as.numeric(lda(type ~ ., dat[[i]][1:j], CV = TRUE)$class)
    out <- ldapredcv
    }, silent=TRUE)
  out 
}

A <- lapply(L, function(x) { apply(x, 2, function(y) y != tr ) })
B <- lapply(A, function(x) { x[,1:10] } )

B <- lapply(B, function(x) { 
                 rownames(x) <- channel[ford]
                 colnames(x) <- paste0('d', 1:dim(x)[2]) 
                 return(x)
    }) 

par(mfrow = c(2,3))
corT <- paste0("Misclassifications ", "F", 0:5)
for(i in 1:6) {
corrplot(B[[i]], method='color', addgrid.col=1, 
         tl.col = ccol[ford],
         mar=3*c(1,0,1,1.5))
title(corT[i])
}
Figure 12: Misclassifications using LDA with CV for various \(d\).

The above plots show when LDA with cross-validation misclassifies each datapoint. A misclassification is denoted by a filled block and columns denote the embedding dimension \(\hat d\).

titlesvor <- paste("LDA decision boundaries for", paste0("F", 0:5))
voronoidf <- lapply(lapply(lda.fit, '[[', 3), data.frame)
#voronoidf <- data.frame(x=lda.fit$means[,1],y=lda.fit$means[,2])

#This creates the voronoi line segments

par(mfrow = c(3,2))
for(i in 1:length(dat)){
  plot(dat[[i]][,1:2], col=ccol3[ford], pch=20, cex=1.5)
  title(titlesvor[i])
  text(dat[[i]][,1:2], labels=rownames(dat[[i]]), 
       pos=ifelse(dat[[i]][,1]<max(dat[[i]][,1] -0.5),4,2), 
       col=ccol3[ford], cex=1.2)

  deldir(x = voronoidf[[i]][,1],y = voronoidf[[i]][,2], rw = c(-15,15,-15,15), 
       plotit=TRUE, add=TRUE, wl='te')
  text(voronoidf[[i]], labels=rownames(voronoidf[[i]]), cex=1.5, pos=1)
}
Figure 13: Voronoi diagrams on class means from LDA on PCA of untransformed correlation matrices

Recall that these are two-dimensional visualizations of a 24-dimensional structure.

pred <- lapply(lapply(lda.fit, predict), '[[', 1)

tabN <- paste0("Table for F", 0:5)
for( i in 1:length(pred)) {
 print(tabN[i])
 print(table(truth,pred = pred[[i]]) )
}
# [1] "Table for F0"
#        pred
# truth   ex in other
#   ex    11  0     0
#   in     0  6     0
#   other  0  0     7
# [1] "Table for F1"
#        pred
# truth   ex in other
#   ex    11  0     0
#   in     0  6     0
#   other  0  0     7
# [1] "Table for F2"
#        pred
# truth   ex in other
#   ex    11  0     0
#   in     0  6     0
#   other  0  0     7
# [1] "Table for F3"
#        pred
# truth   ex in other
#   ex    11  0     0
#   in     0  6     0
#   other  0  0     7
# [1] "Table for F4"
#        pred
# truth   ex in other
#   ex    11  0     0
#   in     0  6     0
#   other  0  0     7
# [1] "Table for F5"
#        pred
# truth   ex in other
#   ex    11  0     0
#   in     0  6     0
#   other  0  0     7
lda.err <- list()
for(i in 1:length(pred)) {
 pr <- as.numeric(pred[[i]])
 lda.err[[i]] <- sum( tr != pr ) / length(tr)
}

Reduce('c', lda.err)
# [1] 0 0 0 0 0 0

The above gives the LDA re-substitution error rates for each of the features.