Search code examples
rggplot2regressionkernel-density

Draw line through 2d density plot

I have a large dataset of gene expression from ~10,000 patient samples (TCGA), and I'm plotting a predicted expression value (x) and the actual observed value (y) of a certain gene signature. For my downstream analysis, I need to draw a precise line through the plot and calculate different parameters in samples above/below the line.
No matter how I draw a line through the data (geom_smooth(method = 'lm', 'glm', 'gam', or 'loess')), the line always seems imperfect - it doesn't cut through the data to my liking (red line is lm in figure).
After playing around for a while, I realized that the 2d kernel density lines (geom_density2d) actually do a good job of showing the slope/trends of my data, so I manually drew a line that kind of cuts through the density lines (black line in figure).

My question: how can I automatically draw a line that cuts through the kernel density lines, as for the black line in the figure? (Rather than manually playing with different intercepts and slopes till something looks good).

The best approach I can think of is to somehow calculate intercept and slope of the longest diameter for each of the kernel lines, take an average of all those intercepts and slopes and plot that line, but that's a bit out of my league. Maybe someone here has experience with this and can help?

A more hacky approach may be getting the x,y coords of each kernel density line from ggplot_build, and going from there, but it feels too hacky (and is also out of my league).

Thanks!

EDIT: Changed a few details to make the figure/analysis easier. (Density lines are smoother now). Reprex:

library(MASS)
set.seed(123)
samples <- 10000
r <- 0.9
data <- mvrnorm(n=samples, mu=c(0, 0), Sigma=matrix(c(2, r, r, 2), nrow=2))
x <- data[, 1]  # standard normal (mu=0, sd=1)
y <- data[, 2]  # standard normal (mu=0, sd=1)

test.df <- data.frame(x = x, y = y)
lm(y ~ x, test.df)

ggplot(test.df, aes(x, y)) +
  geom_point(color = 'grey') +
  geom_density2d(color = 'red', lwd = 0.5, contour = T, h = c(2,2)) + ### EDIT: h = c(2,2)
  geom_smooth(method = "glm", se = F, lwd = 1, color = 'red') +
  geom_abline(intercept = 0, slope = 0.7, lwd = 1, col = 'black') ## EDIT: slope to 0.7

Figure: enter image description here

Solution

  • Tjebo's approach was kind of good initially, but after a close look, I found that it found the longest distance between two points on an ellipse. While this is close to what I wanted, it failed with either an irregular shape of the ellipse, or the sparsity of points in the ellipse. This is because it measured the longest distance between two points; whereas what I really wanted is the longest diameter of an ellipse; i.e.: the semi-major axis. See image below for examples/details.

    Briefly:

    To find/draw density contours of specific density/percentage:
    R - How to find points within specific Contour

    To get the longest diameter ("semi-major axis") of an ellipse: https://stackoverflow.com/a/18278767/3579613

    For function that returns intercept and slope (as in OP), see last piece of code.
    The two pieces of code and images below compare two Tjebo's approach vs. my new approach based on the above posts.

    #### Reprex from OP
require(dplyr)
require(ggplot2)
require(MASS)
set.seed(123)
samples <- 10000
r <- 0.9
data <- mvrnorm(n=samples, mu=c(0, 0), Sigma=matrix(c(2, r, r, 2), nrow=2))
x <- data[, 1]  # standard normal (mu=0, sd=1)
y <- data[, 2]  # standard normal (mu=0, sd=1)
test.df <- data.frame(x = x, y = y)

#### From Tjebo
p <- ggplot(test.df, aes(x, y)) +
  geom_density2d(color = 'red', lwd = 0.5, contour = T, h = 2) 
p_built <- ggplot_build(p)
p_data <- p_built$data[[1]]
p_maxring <- p_data[p_data[['level']] == min(p_data[['level']]),][,2:3]
coord_mean <- c(x = mean(p_maxring$x), y = mean(p_maxring$y))
p_maxring <- p_maxring %>% 
  mutate (mean_dev = sqrt((x - mean(x))^2 + (y - mean(y))^2)) #extra column specifying the distance of each point to the mean of those points
p_maxring = p_maxring[round(seq(1, nrow(p_maxring), nrow(p_maxring)/23)),] #### Make a small ellipse to illustrate flaws of approach
coord_farthest <- c('x' = p_maxring$x[which.max(p_maxring$mean_dev)], 'y' = p_maxring$y[which.max(p_maxring$mean_dev)])
# gives the coordinates of the point farthest away from the mean point
farthest_from_farthest <- sqrt((p_maxring$x - coord_farthest['x'])^2 + (p_maxring$y - coord_farthest['y'])^2)
#now this looks which of the points is the farthest from the point farthest from the mean point :D
coord_fff <- c('x' = p_maxring$x[which.max(farthest_from_farthest)], 'y' = p_maxring$y[which.max(farthest_from_farthest)])
farthest_2_points = data.frame(t(cbind(coord_farthest, coord_fff)))
plot(p_maxring[,1:2], asp=1)
lines(farthest_2_points, col = 'blue', lwd = 2)


#### From answer in another post
d = cbind(p_maxring[,1], p_maxring[,2])
r = ellipsoidhull(d)
exy = predict(r) ## the ellipsoid boundary
lines(exy)
me = colMeans((exy))           
dist2center = sqrt(rowSums((t(t(exy)-me))^2))
max(dist2center)     ## major axis
lines(exy[dist2center == max(dist2center),], col = 'red', lwd = 2)

    enter image description here

    #### The plot here is made from the data in the reprex in OP, but with h = 0.5
library(MASS)
set.seed(123)
samples <- 10000
r <- 0.9
data <- mvrnorm(n=samples, mu=c(0, 0), Sigma=matrix(c(2, r, r, 2), nrow=2))
x <- data[, 1]  # standard normal (mu=0, sd=1)
y <- data[, 2]  # standard normal (mu=0, sd=1)
test.df <- data.frame(x = x, y = y)

## MAKE BLUE LINE
p <- ggplot(test.df, aes(x, y)) +
  geom_density2d(color = 'red', lwd = 0.5, contour = T, h = 0.5)  ## NOTE h = 0.5
p_built <- ggplot_build(p)
p_data <- p_built$data[[1]]
p_maxring <- p_data[p_data[['level']] == min(p_data[['level']]),][,2:3]
coord_mean <- c(x = mean(p_maxring$x), y = mean(p_maxring$y))
p_maxring <- p_maxring %>% 
  mutate (mean_dev = sqrt((x - mean(x))^2 + (y - mean(y))^2))
coord_farthest <- c('x' = p_maxring$x[which.max(p_maxring$mean_dev)], 'y' = p_maxring$y[which.max(p_maxring$mean_dev)])
farthest_from_farthest <- sqrt((p_maxring$x - coord_farthest['x'])^2 + (p_maxring$y - coord_farthest['y'])^2)
coord_fff <- c('x' = p_maxring$x[which.max(farthest_from_farthest)], 'y' = p_maxring$y[which.max(farthest_from_farthest)])

## MAKE RED LINE
## h = 0.5
## Given the highly irregular shape of the contours, I will use only the largest contour line (0.95) for draing the line.
## Thus, average = 1. See function below for details.
ln = long.diam("x", "y", test.df, h = 0.5, average = 1) ## NOTE h = 0.5

## PLOT
ggplot(test.df, aes(x, y)) +
  geom_density2d(color = 'red', lwd = 0.5, contour = T, h = 0.5) + ## NOTE h = 0.5
  geom_segment((aes(x = coord_farthest['x'], y = coord_farthest['y'],
                    xend = coord_fff['x'], yend = coord_fff['y'])), col = 'blue', lwd = 2) +
  geom_abline(intercept = ln[1], slope = ln[2], color = 'red', lwd = 2) +
  coord_equal()

    enter image description here Finally, I came up with the following function to deal with all this. Sorry for the lack of comments/clarity

    #### This will return the intercept and slope of the longest diameter (semi-major axis).
####If Average = TRUE, it will average the int and slope across different density contours.
long.diam = function(x, y, df, probs = c(0.95, 0.5, 0.1), average = T, h = 2) {
  fun.df = data.frame(cbind(df[,x], df[,y]))
  colnames(fun.df) = c("x", "y")
  dens = kde2d(fun.df$x, fun.df$y, n = 200, h = h)
  dx <- diff(dens$x[1:2])
  dy <- diff(dens$y[1:2])
  sz <- sort(dens$z)
  c1 <- cumsum(sz) * dx * dy 
  levels <- sapply(probs, function(x) { 
    approx(c1, sz, xout = 1 - x)$y
  })
  names(levels) = paste0("L", str_sub(formatC(probs, 2, format = 'f'), -2))
  #plot(fun.df$x,fun.df$y, asp = 1)
  #contour(dens, levels = levels, labels=probs, add=T, col = c('red', 'blue', 'green'), lwd = 2)
  #contour(dens, add = T, col = 'red', lwd = 2)
  #abline(lm(fun.df$y~fun.df$x))

  ls <- contourLines(dens, levels = levels)
  names(ls) = names(levels)

  lines.info = list()
  for (i in 1:length(ls)) {
    d = cbind(ls[[i]]$x, ls[[i]]$y)
    exy = predict(ellipsoidhull(d))## the ellipsoid boundary
    colnames(exy) = c("x", "y")
    me = colMeans((exy))            ## center of the ellipse
    dist2center = sqrt(rowSums((t(t(exy)-me))^2))
    #plot(exy,type='l',asp=1)
    #points(d,col='blue')
    #lines(exy[order(dist2center)[1:2],])
    #lines(exy[rev(order(dist2center))[1:2],])
    max.dist = data.frame(exy[rev(order(dist2center))[1:2],])
    line.fit = lm(max.dist$y ~ max.dist$x)
    lines.info[[i]] = c(as.numeric(line.fit$coefficients[1]), as.numeric(line.fit$coefficients[2]))
  }
  names(lines.info) = names(ls)

  #plot(fun.df$x,fun.df$y, asp = 1)
  #contour(dens, levels = levels, labels=probs, add=T, col = c('red', 'blue', 'green'), lwd = 2)
  #abline(lines.info[[1]], col = 'red', lwd = 2)
  #abline(lines.info[[2]], col = 'blue', lwd = 2)
  #abline(lines.info[[3]], col = 'green', lwd = 2)
  #abline(apply(simplify2array(lines.info), 1, mean), col = 'black', lwd = 4)
  if (isTRUE(average)) {
    apply(simplify2array(lines.info), 1, mean)
  } else {
    lines.info[[average]]
  }
}

    Finally, here's the final implementation of the different answers:

    library(MASS)
set.seed(123)
samples = 10000
r = 0.9
data = mvrnorm(n=samples, mu=c(0, 0), Sigma=matrix(c(2, r, r, 2), nrow=2))
x = data[, 1]  # standard normal (mu=0, sd=1)
y = data[, 2]  # standard normal (mu=0, sd=1)
#plot(x, y)
test.df = data.frame(x = x, y = y)

#### Find furthest two points of contour
## BLUE
p <- ggplot(test.df, aes(x, y)) +
  geom_density2d(color = 'red', lwd = 2, contour = T, h = 2) 
p_built <- ggplot_build(p)
p_data <- p_built$data[[1]]
p_maxring <- p_data[p_data[['level']] == min(p_data[['level']]),][,2:3]
coord_mean <- c(x = mean(p_maxring$x), y = mean(p_maxring$y))
p_maxring <- p_maxring %>% 
  mutate (mean_dev = sqrt((x - mean(x))^2 + (y - mean(y))^2))
coord_farthest <- c('x' = p_maxring$x[which.max(p_maxring$mean_dev)], 'y' = p_maxring$y[which.max(p_maxring$mean_dev)])
farthest_from_farthest <- sqrt((p_maxring$x - coord_farthest['x'])^2 + (p_maxring$y - coord_farthest['y'])^2)
coord_fff <- c('x' = p_maxring$x[which.max(farthest_from_farthest)], 'y' = p_maxring$y[which.max(farthest_from_farthest)])

#### Find the average intercept and slope of 3 contour lines (0.95, 0.5, 0.1), as in my long.diam function above.
## RED
ln = long.diam("x", "y", test.df)

#### Plot everything. Black line is GLM
ggplot(test.df, aes(x, y)) +
  geom_point(color = 'grey') +
  geom_density2d(color = 'red', lwd = 1, contour = T, h = 2) + 
  geom_smooth(method = "glm", se = F, lwd = 1, color = 'black') +
  geom_abline(intercept = ln[1], slope = ln[2], col = 'red', lwd = 1) +
  geom_segment((aes(x = coord_farthest['x'], y = coord_farthest['y'],
                    xend = coord_fff['x'], yend = coord_fff['y'])), col = 'blue', lwd = 1) +
  coord_equal()

    Final image