test_downcentric_whatever.R

imputations.R

library("impute")

# simulating a matrix
peptideCnt = 1000
runsCnt = 6

expressions = matrix(
  data=rnorm(peptideCnt*runsCnt, mean=1000, sd=400),
  nrow=peptideCnt,
  ncol=runsCnt
)

draw_missing_at_random_mask = function(expressions, prob = c(0.75, 0.25)) sample(
  x = c(FALSE, TRUE),
  size = nrow(expressions) * ncol(expressions),
  prob = prob,
  replace = TRUE
)

missing_at_random_mask = draw_missing_at_random_mask(expressions) 
table(missing_at_random_mask)

expressions_with_mocked_NAs = expressions
expressions_with_mocked_NAs[missing_at_random_mask] = NA

expressions_with_mocked_NAs_imputed = impute.knn(expressions_with_mocked_NAs, k=10)$data

nrow(expressions_with_mocked_NAs_imputed)
ncol(expressions_with_mocked_NAs_imputed)

# values not needing imputation untouched? YES
all(expressions_with_mocked_NAs_imputed[!missing_at_random_mask] == expressions[!missing_at_random_mask])

# errors:
err = expressions_with_mocked_NAs_imputed[missing_at_random_mask] - expressions[missing_at_random_mask]
hist(err, n=100)



# then: comparison of 2 sample conditions.
cond_A = matrix(
  data=rnorm(peptideCnt*runsCnt, mean=1000, sd=100),
  nrow=peptideCnt,
  ncol=runsCnt
)
cond_B = matrix(
  data=rnorm(peptideCnt*runsCnt, mean=1200, sd=100),
  nrow=peptideCnt,
  ncol=runsCnt
)

# do t.tests
do_tests = function(cond_A, cond_B){
  if( nrow(cond_A) == nrow(cond_B) ){
    pvalues = numeric(nrow(cond_A))
    for( i in 1:nrow(cond_A)){
      test_result = t.test( cond_A[i,], cond_B[i,] )
      pvalues[i] = test_result$p.value
    }
  }
  pvalues
}

p_values = do_tests(cond_A, cond_B)
adj_p_values = p.adjust(p_values, method="fdr")
plot(p_values[p_values< 0.1], adj_p_values[p_values< 0.1], pch=".")
abline(coef = c(0,1))
abline(h=0.01)
abline(v=0.01)

all_discoveries_mask = adj_p_values < 0.01


cond_A_mocked_NAs = cond_A
cond_A_mock_mask = draw_missing_at_random_mask(cond_A)
cond_A_mocked_NAs[cond_A_mock_mask] = NA
cond_A_mocked_NAs_imputed = impute.knn(cond_A_mocked_NAs, k=10)$data

cond_B_mocked_NAs = cond_B
cond_B_mock_mask = draw_missing_at_random_mask(cond_B)
cond_B_mocked_NAs[cond_B_mock_mask] = NA
cond_B_mocked_NAs_imputed = impute.knn(cond_B_mocked_NAs, k=10)$data

imputed_p_values = do_tests(cond_A_mocked_NAs_imputed, cond_B_mocked_NAs_imputed)
adj_imputed_p_values = p.adjust(imputed_p_values, method="fdr")

all_imputed_discoveries_mask = adj_imputed_p_values < 0.01

table(all_discoveries_mask, all_imputed_discoveries_mask)

# should the values be imputed within groups? Which groups? Likely technical replicates.

# what is t.test doing with missing values as such?
?t.test
getOption("na.action") # so simply omits them