Marginal Distribution of \(z\) Scores: Alternative

Introduction

We’ve seen that when generated from the global null, that is, when the cases and controls have no difference, the \(z\) scores’ behavior are what one would expect if simulated from correlated marginally \(N\left(0, 1\right)\) random variables.

But what if the \(z\) scores are not simulated from the global null? Are these \(z\) scores going to behave significantly different from correlated marginally \(N\left(0, 1\right)\) random samples? Let’s take a look at the real data containing true effects.

library(limma)
library(edgeR)
library(qvalue)
library(ashr)

#extract top g genes from G by n matrix X of expression

top_genes_index = function (g, X)
{return(order(rowSums(X), decreasing = TRUE)[1 : g])
}

lcpm = function (r) {
  R = colSums(r)
  t(log2(((t(r) + 0.5) / (R + 1)) * 10^6))
}

# transform counts to z scores
# these z scores are marginally N(0, 1) under null

counts_to_z = function (counts, condition) {
  design = model.matrix(~condition)
  dgecounts = calcNormFactors(DGEList(counts = counts, group = condition))
  v = voom(dgecounts, design, plot = FALSE)
  lim = lmFit(v)
  r.ebayes = eBayes(lim)
  p = r.ebayes$p.value[, 2]
  t = r.ebayes$t[, 2]
  z = sign(t) * qnorm(1 - p/2)
  return (z)
}

Generating non-null \(z\) scores from real data

For convenience and without loss of generality, we are using the liver tissue as an anchor in the simulation, and always choose top expressed genes in livers.

r.liver = read.csv("../data/liver.csv")
r.liver = r.liver[, -(1 : 2)] # remove gene name and description
Y = lcpm(r.liver)
G = 1e4
subset = top_genes_index(G, Y)
r.liver = r.liver[subset, ]

tissue = "heart"
r = read.csv(paste0("../data/", tissue, ".csv"))
r = r[, -(1 : 2)] # remove gene name and description
## choose top expressed genes in liver
r = r[subset, ]

set.seed(777)
m = 1e3
n = 5
z.list = list()
condition = c(rep(0, n), rep(1, n))
for (i in 1 : m) {
  counts = cbind(r.liver[, sample(1 : ncol(r.liver), n)],
                 r[, sample(1 : ncol(r), n)])
  z.list[[i]] = counts_to_z(counts, condition)
}

z.mat = matrix(unlist(z.list), nrow = m, byrow = TRUE)
saveRDS(z.mat, "../output/z_5liver_5heart_777.rds")

tissue = "muscle"
r = read.csv(paste0("../data/", tissue, ".csv"))
r = r[, -(1 : 2)] # remove gene name and description
## choose top expressed genes in liver
r = r[subset, ]

set.seed(777)
m = 1e3
n = 5
z.list = list()
condition = c(rep(0, n), rep(1, n))
for (i in 1 : m) {
  counts = cbind(r.liver[, sample(1 : ncol(r.liver), n)],
                 r[, sample(1 : ncol(r), n)])
  z.list[[i]] = counts_to_z(counts, condition)
}

z.mat = matrix(unlist(z.list), nrow = m, byrow = TRUE)
saveRDS(z.mat, "../output/z_5liver_5muscle_777.rds")

Marginal distributions of correlated non-null \(z\) scores

Now following the simulations to explore the marginal distributions of the correlated null \(z\) scores, we are taking a look at the marginal distributions of the correlated non-null \(z\) scores. The main focus is still the number of tail observations.

Liver vs Heart

z.heart = readRDS("../output/z_5liver_5heart_777.rds")
n = ncol(z.heart)
m = nrow(z.heart)

Liver vs Muscle

z.muscle = readRDS("../output/z_5liver_5muscle_777.rds")
n = ncol(z.muscle)
m = nrow(z.muscle)

Conclusion

As expected, the empirical distribution and the indicated marginal distribution of the correlated non-null \(z\) scores are starkly different from those of the correlated null ones.

Particular interesting are the column-wise plots. The distribution of the number of tail observations is not only not close to normal, not close to what would be expected under correlated marginally \(N\left(0, 1\right)\), but also not even unimodal, not even peaked at \(m\alpha\) when \(\alpha \leq0.5\) or \(m\left(1-\alpha\right)\) when \(\alpha\geq0.5\). It suggests that there are indeed plenty of true signals, which make the marginal distribution of the \(z\) scores for a certain gene often not centered at \(0\).

Session information

sessionInfo()

R version 3.3.3 (2017-03-06)
Platform: x86_64-apple-darwin13.4.0 (64-bit)
Running under: macOS Sierra 10.12.5

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] ashr_2.1-13   qvalue_2.4.2  edgeR_3.14.0  limma_3.28.21

loaded via a namespace (and not attached):
 [1] Rcpp_0.12.10      knitr_1.16        magrittr_1.5     
 [4] MASS_7.3-47       splines_3.3.3     pscl_1.4.9       
 [7] doParallel_1.0.10 munsell_0.4.3     lattice_0.20-35  
[10] SQUAREM_2016.10-1 colorspace_1.3-2  rlang_0.1.1      
[13] foreach_1.4.3     stringr_1.2.0     plyr_1.8.4       
[16] tools_3.3.3       parallel_3.3.3    grid_3.3.3       
[19] gtable_0.2.0      git2r_0.18.0      iterators_1.0.8  
[22] htmltools_0.3.6   yaml_2.1.14       lazyeval_0.2.0   
[25] rprojroot_1.2     digest_0.6.12     tibble_1.3.1     
[28] Matrix_1.2-10     reshape2_1.4.2    ggplot2_2.2.1    
[31] codetools_0.2-15  evaluate_0.10     rmarkdown_1.5    
[34] stringi_1.1.5     scales_0.4.1      backports_1.0.5  
[37] truncnorm_1.0-7