558 lines
17 KiB
R
558 lines
17 KiB
R
#########################################################
|
|
# TASK: To compare OR from master snps
|
|
# chisq, fisher test and logistic and adjusted logistic
|
|
#########################################################
|
|
getwd()
|
|
setwd('~/git/LSHTM_analysis/scripts')
|
|
getwd()
|
|
|
|
#install.packages("logistf")
|
|
library(logistf)
|
|
#########################################################
|
|
#%% variable assignment: input and output paths & filenames
|
|
drug = 'pyrazinamide'
|
|
gene = 'pncA'
|
|
gene_match = paste0(gene,'_p.')
|
|
cat(gene_match)
|
|
|
|
#===========
|
|
# input and output dirs
|
|
#===========
|
|
datadir = paste0('~/git/Data')
|
|
indir = paste0(datadir, '/', drug, '/', 'input')
|
|
outdir = paste0(datadir, '/', drug, '/', 'output')
|
|
|
|
#===========
|
|
# input and output files
|
|
#===========
|
|
in_filename = 'original_tanushree_data_v2.csv'
|
|
#in_filename = 'mtb_gwas_v3.csv'
|
|
infile = paste0(datadir, '/', in_filename)
|
|
cat(paste0('Reading infile1: raw snps', ' ', infile) )
|
|
|
|
# infile2: _gene associated meta snps file to extract valid snps and add calcs to.
|
|
# This is outfile3 from snps_extraction.py
|
|
in_filename_metadata = paste0(tolower(gene), '_metadata.csv')
|
|
infile_metadata = paste0(outdir, '/', in_filename_metadata)
|
|
cat(paste0('Reading infile2: gene associated metadata:', infile_metadata))
|
|
|
|
#===========
|
|
# output
|
|
#===========
|
|
out_filename = paste0(tolower(gene),'_', 'af_or.csv')
|
|
outfile = paste0(outdir, '/', out_filename)
|
|
cat(paste0('Output file with full path:', outfile))
|
|
#%% end of variable assignment for input and output files
|
|
|
|
#########################################################
|
|
# 1: Read master/raw snps stored in snps/
|
|
#####################################################
|
|
|
|
#===============
|
|
# Step 1: read raw snps (all remove entries with NA in pza column)
|
|
#===============
|
|
raw_data_all = read.csv(infile, stringsAsFactors = F)
|
|
|
|
# building cols to extract
|
|
dr_muts_col = paste0('dr_mutations_', drug)
|
|
other_muts_col = paste0('other_mutations_', drug)
|
|
|
|
cat('Extracting columns based on variables:\n'
|
|
, drug
|
|
, '\n'
|
|
, dr_muts_col
|
|
, '\n'
|
|
, other_muts_col
|
|
, '\n===============================================================')
|
|
|
|
raw_data = raw_data_all[,c("id"
|
|
, drug
|
|
, dr_muts_col
|
|
, other_muts_col)]
|
|
rm(raw_data_all)
|
|
|
|
rm(indir, in_filename, infile)
|
|
|
|
#===========
|
|
# 1a: exclude na
|
|
#===========
|
|
raw_data = raw_data[!is.na(raw_data[[drug]]),]
|
|
|
|
total_samples = length(unique(raw_data$id))
|
|
cat(paste0('Total samples without NA in', ' ', drug, 'is:', total_samples))
|
|
|
|
# sanity check: should be true
|
|
is.numeric(total_samples)
|
|
|
|
#===========
|
|
# 1b: combine the two mutation columns
|
|
#===========
|
|
#raw_data$all_mutations_pyrazinamide = paste(raw_data$dr_mutations_pyrazinamide, raw_data$other_mutations_pyrazinamide)
|
|
|
|
all_muts_colname = paste0('all_mutations_', drug)
|
|
raw_data[[all_muts_colname]] = paste(raw_data[[dr_muts_col]], raw_data[[other_muts_col]])
|
|
head(raw_data[[all_muts_colname]])
|
|
|
|
#===========
|
|
# 1c: create yet another column that contains all the mutations but in lower case
|
|
#===========
|
|
head(raw_data[[all_muts_colname]])
|
|
raw_data$all_muts_gene = tolower(raw_data[[all_muts_colname]])
|
|
head(raw_data$all_muts_gene)
|
|
|
|
# sanity checks
|
|
#table(grepl("gene_p",raw_data$all_muts_gene))
|
|
cat(paste0('converting gene match:', gene_match, ' ', 'to lowercase'))
|
|
gene_match = tolower(gene_match)
|
|
|
|
table(grepl(gene_match,raw_data$all_muts_gene))
|
|
|
|
# sanity check
|
|
if(sum(table(grepl(gene_match, raw_data$all_muts_gene))) == total_samples){
|
|
cat('PASS: Total no. of samples match')
|
|
} else{
|
|
cat('FAIL: No. of samples mismatch')
|
|
}
|
|
|
|
#########################################################
|
|
# 2: Read valid snps for which OR
|
|
# can be calculated
|
|
#########################################################
|
|
cat(paste0('Reading metadata infile:', infile_metadata))
|
|
|
|
gene_metadata = read.csv(infile_metadata
|
|
#, file.choose()
|
|
, stringsAsFactors = F
|
|
, header = T)
|
|
|
|
|
|
# clear variables
|
|
rm(in_filename_metadata, infile_metadata)
|
|
|
|
# count na in pyrazinamide column
|
|
tot_pza_na = sum(is.na(gene_metadata$pyrazinamide))
|
|
expected_rows = nrow(gene_metadata) - tot_pza_na
|
|
|
|
# drop na from the pyrazinamide colum
|
|
gene_snps_or = gene_metadata[!is.na(gene_metadata[[drug]]),]
|
|
|
|
# sanity check
|
|
if(nrow(gene_snps_or) == expected_rows){
|
|
cat('PASS: no. of rows match with expected_rows')
|
|
} else{
|
|
cat('FAIL: nrows mismatch.')
|
|
}
|
|
|
|
# extract unique snps to iterate over for AF and OR calcs
|
|
gene_snps_unique = unique(gene_snps_or$mutation)
|
|
|
|
cat(paste0('Total no. of distinct comp snps to perform OR calcs: ', length(gene_snps_unique)))
|
|
#=====================================
|
|
#OR calcs using the following 4
|
|
#1) chisq.test
|
|
#2) fisher
|
|
#3) modified chisq.test
|
|
#4) logistic
|
|
#5) adjusted logistic?
|
|
|
|
#======================================
|
|
#########################
|
|
# custom chisq function:
|
|
# To calculate OR
|
|
#########################
|
|
#i = "pnca_p.trp68gly"
|
|
#mut = grepl(i,raw_data$all_muts_gene)
|
|
#dst = raw_data[[drug]]
|
|
|
|
#x = as.numeric(mut)
|
|
#y = dst
|
|
|
|
custom_chisq_or = function(x,y){
|
|
tab = as.matrix(table(x,y))
|
|
a = tab[2,2]
|
|
if (a==0){ a<-0.5}
|
|
b = tab[2,1]
|
|
if (b==0){ b<-0.5}
|
|
c = tab[1,2]
|
|
if (c==0){ c<-0.5}
|
|
d = tab[1,1]
|
|
if (d==0){ d<-0.5}
|
|
(a/b)/(c/d)
|
|
|
|
}
|
|
|
|
#======================================
|
|
#==============
|
|
# TEST WITH ONE
|
|
#==============
|
|
i = "pnca_p.trp68gly"
|
|
i = "pnca_p.gln10pro"
|
|
i = "pnca_p.leu159arg"
|
|
|
|
# IV
|
|
table(grepl(i,raw_data$all_muts_gene))
|
|
mut = grepl(i,raw_data$all_muts_gene)
|
|
|
|
# DV
|
|
#dst = raw_data$pyrazinamide
|
|
dst = raw_data[[drug]] # or raw_data[,drug]
|
|
|
|
# 2X2 table
|
|
table(mut, dst)
|
|
|
|
# CV
|
|
#c = raw_data$id[mut]
|
|
c = raw_data$id[grepl(i,raw_data$all_muts_gene)]
|
|
#sid = grepl(raw_data$id[mut], raw_data$id) # warning
|
|
#argument 'pattern' has length > 1 and only the first element will be used
|
|
#grepl(raw_data$id=="ERR2512440", raw_data$id)
|
|
|
|
sid = grepl(paste(c,collapse="|"), raw_data$id)
|
|
table(sid)
|
|
|
|
# 3X2 table
|
|
table(mut, dst, sid)
|
|
|
|
#===================================================
|
|
# compare ORs from different calcs
|
|
|
|
#1) chisq
|
|
chisq.test(table(mut,dst))
|
|
chisq_estimate = chisq.test(table(mut,dst))$statistic
|
|
est_chisq = chisq_estimate[[1]]; print(paste0('chi sq estimate:', est_chisq))# numeric part only
|
|
pval_chisq = chisq.test(table(mut,dst))$p.value; print(paste0('pvalue:', pval_chisq))
|
|
|
|
|
|
#2) fisher
|
|
fisher.test(table(mut, dst))
|
|
fisher.test(table(mut, dst))$p.value
|
|
est_fisher = fisher.test(table(mut, dst))$estimate
|
|
or_fisher = est_fisher[[1]]; print(paste0('OR fisher:', or_fisher))# numeric part only
|
|
pval_fisher = fisher.test(table(mut, dst))$p.value; print(paste0('pval fisher:', pval_fisher))
|
|
|
|
|
|
#3) custom chisq
|
|
or_mychisq = custom_chisq_or(mut,dst)
|
|
|
|
#4) logistic
|
|
summary(model<-glm(dst ~ mut, family = binomial))
|
|
|
|
or_logistic = exp(summary(model)$coefficients[2,1]); print(paste0('OR logistic:', or_logistic))
|
|
pval_logistic = summary(model)$coefficients[2,4]; print(paste0('pval logistic:', pval_logistic))
|
|
|
|
# extract SE of the logistic model for a given snp
|
|
logistic_se = summary(model)$coefficients[2,2]; print(paste0('SE:', logistic_se))
|
|
|
|
# extract Z of the logistic model for a given snp
|
|
logistic_zval = summary(model)$coefficients[2,3]; print(paste0('Z-value:', logistic_zval))
|
|
|
|
# extract confint interval of snp (2 steps, since the output is a named number)
|
|
ci_mod = exp(confint(model))[2,]; print(paste0('CI:', ci_mod))
|
|
#logistic_ci = paste(ci_mod[["2.5 %"]], ",", ci_mod[["97.5 %"]])
|
|
|
|
logistic_ci_lower = ci_mod[["2.5 %"]]; print(paste0('CI_lower:', logistic_ci_lower))
|
|
logistic_ci_upper = ci_mod[["97.5 %"]]; print(paste0('CI_upper:', logistic_ci_upper))
|
|
|
|
# adjusted logistic or: doesn't seem to make a difference
|
|
summary(model2<-glm(dst ~ mut + sid, family = binomial))
|
|
or_logistic2 = exp(summary(model2)$coefficients[2,1]); print(paste0('Adjusted OR logistic:', or_logistic2))
|
|
pval_logistic2 = summary(model2)$coefficients[2,4]; print(paste0('Adjusted pval logistic:',pval_logistic2))
|
|
|
|
# print all output
|
|
writeLines(c(paste0("mutation:", i)
|
|
, paste0("=========================")
|
|
, paste0("OR logistic:", or_logistic,"--->", "pval logistic:", pval_logistic )
|
|
, paste0("OR adjusted logistic:", or_logistic2,"--->", "pval adjusted logistic:", pval_logistic2)
|
|
, paste0("OR fisher:", or_fisher, "--->","pval fisher:", pval_fisher )
|
|
, paste0("OR custom chisq:", or_mychisq, "--->","pval fisher:", pval_fisher )
|
|
, paste0("Chisq estimate:", est_chisq, "--->","pval chisq:", pval_chisq)))
|
|
#%%========================================================
|
|
# looping with sapply: a subset of mutations
|
|
#%%========================================================
|
|
|
|
snps = c("pnca_p.trp68gly", "pnca_p.leu4ser", "pnca_p.leu159arg","pnca_p.his57arg" )
|
|
#snps = snps_test[1:2]
|
|
snps
|
|
|
|
# custom chisq
|
|
ors = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
custom_chisq_or(mut,dst)
|
|
})
|
|
head(ors)
|
|
|
|
# pvalue fisher, to be used with custom chisq
|
|
pvals = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
fisher.test(mut,dst)$p.value
|
|
})
|
|
head(pvals)
|
|
|
|
# allele frequency
|
|
afs = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
mean(mut)
|
|
})
|
|
head(afs)
|
|
|
|
# logistic reg parameters: individual sapply
|
|
|
|
#--------------
|
|
## logistci or
|
|
#--------------
|
|
ors_logistic = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut, family = binomial)
|
|
or_logistic = exp(summary(model)$coefficients[2,1])
|
|
|
|
})
|
|
ors_logistic
|
|
head(ors_logistic); head(names(ors_logistic))
|
|
|
|
#-------------------
|
|
## logistic p-value
|
|
#--------------
|
|
pvals_logistic = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut , family = binomial)
|
|
pval_logistic = summary(model)$coefficients[2,4]
|
|
})
|
|
|
|
head(pvals_logistic); head(names(pvals_logistic))
|
|
|
|
#--------------
|
|
## logistic se
|
|
#--------------
|
|
se_logistic = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut , family = binomial)
|
|
logistic_se = summary(model)$coefficients[2,2]
|
|
})
|
|
|
|
head(se_logistic); head(names(se_logistic))
|
|
|
|
#--------------
|
|
## logistic z-value
|
|
#--------------
|
|
zval_logistic = sapply(gene_snps_unique,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut , family = binomial)
|
|
logistic_zval = summary(model)$coefficients[2,3]
|
|
})
|
|
|
|
head(zval_logistic); head(names(zval_logistic))
|
|
#--------------
|
|
## logistic ci - lower bound
|
|
#--------------
|
|
ci_lb_logistic = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut , family = binomial)
|
|
ci_mod = exp(confint(model))[2,]
|
|
logistic_ci_lower = ci_mod[["2.5 %"]]
|
|
})
|
|
|
|
head(ci_lb_logistic); head(names(ci_lb_logistic))
|
|
#--------------
|
|
## logistic ci - upper bound
|
|
#--------------
|
|
ci_ub_logistic = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut , family = binomial)
|
|
ci_mod = exp(confint(model))[2,]
|
|
logistic_ci_upper = ci_mod[["97.5 %"]]
|
|
|
|
})
|
|
|
|
head(ci_ub_logistic); head(names(ci_ub_logistic))
|
|
|
|
#--------------
|
|
# adjusted logistic with sample id: Doesn't seem to make a difference
|
|
#--------------
|
|
logistic_ors2 = sapply(snps,function(m){
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
c = raw_data$id[mut]
|
|
sid = grepl(paste(c,collapse="|"), raw_data$id)
|
|
model2<-glm(dst ~ mut + sid, family = binomial)
|
|
or_logistic2 = exp(summary(model2)$coefficients[2,1])
|
|
#pval_logistic2 = summary(model2)$coefficients[2,4]
|
|
})
|
|
head(logistic_ors)
|
|
|
|
#=!=!=!=!=!=!=!=!=!=!=!=!=!=!
|
|
# COMMENT: individual sapply seem wasteful
|
|
#=!=!=!=!=!=!=!=!=!=!=!=!=!=!
|
|
|
|
#%%========================================================
|
|
# sapply with multiple values being returned as df
|
|
#Link: https://gist.github.com/primaryobjects/33adabc337edd67b4a8d
|
|
#%%========================================================
|
|
snps = c("pnca_p.trp68gly", "pnca_p.leu4ser", "pnca_p.leu159arg","pnca_p.his57arg" )
|
|
#snps = snps_test[1:4]
|
|
snps
|
|
|
|
# yayy works!
|
|
# DV: pyrazinamide 0 or 1
|
|
dst = raw_data[[drug]]
|
|
|
|
# initialise an empty df
|
|
or_df = data.frame()
|
|
|
|
x = sapply(snps,function(m){
|
|
|
|
#df = data.frame()
|
|
|
|
mut = grepl(m,raw_data$all_muts_gene)
|
|
model<-glm(dst ~ mut, family = binomial)
|
|
|
|
# allele frequency
|
|
afs = mean(mut)
|
|
|
|
# logistic model
|
|
beta_logistic = summary(model)$coefficients[2,1]
|
|
|
|
or_logistic = exp(summary(model)$coefficients[2,1])
|
|
print(paste0('logistic OR:', or_logistic))
|
|
|
|
pval_logistic = summary(model)$coefficients[2,4]
|
|
print(paste0('logistic pval:', pval_logistic))
|
|
|
|
se_logistic = summary(model)$coefficients[2,2]
|
|
zval_logistic = summary(model)$coefficients[2,3]
|
|
ci_mod = exp(confint(model))[2,]
|
|
ci_lower_logistic = ci_mod[["2.5 %"]]
|
|
ci_upper_logistic = ci_mod[["97.5 %"]]
|
|
|
|
# custom_chisq and fisher: OR p-value and CI
|
|
or_mychisq = custom_chisq_or(dst, mut)
|
|
|
|
or_fisher = fisher.test(dst, mut)$estimate
|
|
or_fisher = or_fisher[[1]]
|
|
|
|
pval_fisher = fisher.test(dst, mut)$p.value
|
|
|
|
ci_lower_fisher = fisher.test(dst, mut)$conf.int[1]
|
|
ci_upper_fisher = fisher.test(dst, mut)$conf.int[2]
|
|
|
|
# chi sq estimates
|
|
estimate_chisq = chisq.test(dst, mut)$statistic; estimate_chisq
|
|
est_chisq = estimate_chisq[[1]]; print(est_chisq)
|
|
|
|
pval_chisq = chisq.test(dst, mut)$p.value
|
|
|
|
# build a row to append to df
|
|
row = data.frame(mutation = m
|
|
, af = afs
|
|
, beta_logistic = beta_logistic
|
|
, or_logistic = or_logistic
|
|
, pval_logistic = pval_logistic
|
|
, se_logistic = se_logistic
|
|
, zval_logistic = zval_logistic
|
|
, ci_low_logistic = ci_lower_logistic
|
|
, ci_hi_logistic = ci_upper_logistic
|
|
, or_mychisq = or_mychisq
|
|
, or_fisher = or_fisher
|
|
, pval_fisher = pval_fisher
|
|
, ci_low_fisher= ci_lower_fisher
|
|
, ci_hi_fisher = ci_upper_fisher
|
|
, est_chisq = est_chisq
|
|
, pval_chisq = pval_chisq
|
|
)
|
|
#print(row)
|
|
|
|
or_df <<- rbind(or_df, row)
|
|
|
|
})
|
|
|
|
write.csv(or_df, 'test_ors.csv')
|
|
|
|
#=================================
|
|
# testing logistic or with maxit, etc.
|
|
|
|
print(paste0('subset to iterate over;', snps))
|
|
|
|
# start loop
|
|
perfectSeparation <- function(w) {
|
|
if(grepl("fitted probabilities numerically 0 or 1 occurred",
|
|
as.character(w))) {ww <<- ww+1}
|
|
}
|
|
|
|
for (i in snps){
|
|
|
|
print(i)
|
|
|
|
# IV
|
|
#mut<-as.numeric(grepl(i,raw_data$all_muts_gene))
|
|
mut = grepl(i,raw_data$all_muts_gene)
|
|
table(mut)
|
|
|
|
# DV
|
|
#dst<-as.numeric(raw_data[[drug]])
|
|
dst = raw_data[[drug]]
|
|
|
|
# table
|
|
print(table(dst, mut))
|
|
|
|
#=====================
|
|
# logistic regression, glm.control(maxit = n)
|
|
#https://stats.stackexchange.com/questions/11109/how-to-deal-with-perfect-separation-in-logistic-regression
|
|
#=====================
|
|
#n = 1
|
|
summary(model<-glm(dst ~ mut
|
|
, family = binomial
|
|
#, control = glm.control(maxit = 1)
|
|
#, options(warn = 1)
|
|
))
|
|
#, warning = perfectSeparation))
|
|
or_logistic = exp(summary(model)$coefficients[2,1]); print(or_logistic)
|
|
pval_logistic_maxit = summary(model)$coefficients[2,4]; print(pval_logistic_maxit)
|
|
logistic_se = summary(model)$coefficients[2,2]
|
|
logistic_zval = summary(model)$coefficients[2,3]
|
|
ci_mod = exp(confint(model))[2,]
|
|
logistic_ci_lower = ci_mod[["2.5 %"]]
|
|
logistic_ci_upper = ci_mod[["97.5 %"]]
|
|
|
|
#=====================
|
|
# fishers test
|
|
#=====================
|
|
#attributes(fisher.test(table(dst, mut)))
|
|
or_fisher = fisher.test(table(dst, mut))$estimate
|
|
or_fisher = or_fisher[[1]]; or_fisher
|
|
|
|
pval_fisher = fisher.test(table(dst, mut))$p.value ; pval_fisher
|
|
|
|
#=====================
|
|
# custom chi square
|
|
#=====================
|
|
or_mychisq = custom_chisq_or(mut,dst)
|
|
|
|
#=====================
|
|
# chi square
|
|
#=====================
|
|
#chisq.test(y = dst, x = mut)
|
|
#attributes(chisq.test(table(dst, mut)))
|
|
est_chisq = chisq.test(table(dst, mut))$statistic
|
|
est_chisq = est_chisq[[1]]; est_chisq
|
|
|
|
pval_chisq = chisq.test(table(dst, mut))$p.value; pval_chisq
|
|
|
|
# all output
|
|
writeLines(c(paste0("mutation:", i)
|
|
, paste0("=========================")
|
|
, paste0("OR logistic:", or_logistic,"--->", "pval logistic:", pval_logistic )
|
|
, paste0("OR fisher:", or_fisher, "--->","pval fisher:", pval_fisher )
|
|
, paste0("OR custom chisq:", or_mychisq, "--->","pval fisher:", pval_fisher )
|
|
, paste0("Chisq estimate:", est_chisq, "--->","pval chisq:", pval_chisq)))
|
|
}
|
|
#=====================
|
|
# fishers test
|
|
#=====================
|
|
#attributes(fisher.test(table(dst, mut)))
|
|
or_fisher = fisher.test(table(dst, mut))$estimate
|
|
or_fisher = or_fisher[[1]]; or_fisher
|
|
|
|
pval_fisher = fisher.test(table(dst, mut))$p.value ; pval_fisher
|
|
|
|
|
|
# https://stats.stackexchange.com/questions/259635/what-is-the-difference-using-a-fishers-exact-test-vs-a-logistic-regression-for
|
|
exact2x2(table(dst, mut),tsmethod="central")
|
|
#=====================================================================
|