######################################################################
#                   Factorial analysis code
######################################################################
#
# Version 1 by Sohela Shah and Saunak Sen
# Date 11 September 2009
#
#######################################################################
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# any later version.
#
# This program is distributed in the hope that it will be useful, but
# WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
# General Public License for more details.
#
# A copy of the GNU General Public License is available at
# http://www.gnu.org/licenses/.
#######################################################################
# This code is meant to accompany the paper entitled Strain Background 
# Modifies Phenotypes in the Atp8b1-Deficient Mouse 
# by S. Shah, U.R. Sanford, J. Vargas, H. Xu, A. Groen, 
# C.C. Paulusma, L. Pawlikowska, S. Sen, R.P.J. Oude Elferink, 
# L.N. Bull
########################################################################
# Our objective is to analyze the strain data to discover differences
# between strains (C57BL/6, 129S, and F1), genotype (mutant or wild
# type), the diet regime (control or cholate), and sex.  The initial
# (base) analysis focused on these factors and their interaction. Two
# 129S sub-strains (129S1 and 129S4) were used. These were combined
# into one group for both pure strains and F1 mice.
#
# The main phenotypes assayed are as follows: Weight loss,
# serum cholesterol, ALP, bilirubin, liver weight, bile biochemisty (bile
# cholesterol, bile phospholipids, and bile salts). The serums
# biochemistry levels were assayed at start of experiment, and on day
# of sacrifice. The bile biochemistry levels were assayed at
# sacrifice. Most phenotypes were log transformed using natural
# logarithms.
#
# Enter data from four sheets saved as CSV files in MS Excel and then
# combine them into a single table. The column names were consistent
# across sheets, which was checked. Two columns were added for bile
# cholesterol and bile phospholipids. The script file was changed
# accordingly to reflect this change. Rows with no data were deleted.

# input data files

wt.control <- read.delim("/Users/sohelashah/Documents/Cholestasis/Mouse data/R files/WT on Control.txt")
wt.cholate <- read.delim("/Users/sohelashah/Documents/Cholestasis/Mouse data/R files/WT on Cholate.txt")
mut.control <- read.delim("/Users/sohelashah/Documents/Cholestasis/Mouse data/R files/Mutant on Control.txt")
mut.cholate <- read.delim("/Users/sohelashah/Documents/Cholestasis/Mouse data/R files/Mutant on Cholate.txt")

# keep subset to rows with meaningful data, remove rows with no data

idx <- complete.cases(wt.control[,2])
wt.control <-wt.control[idx,]
idx <- complete.cases(wt.cholate[,2])
wt.cholate <-wt.cholate[idx,]
idx <- complete.cases(mut.control[,2])
mut.control <-mut.control[idx,]
idx <- complete.cases(mut.cholate[,2])
mut.cholate <-mut.cholate[idx,]

idx.c <- c(2,5,6,11,16:20,23:27,37:38,45,12,15,21,28,22,29,34:36,46:63,10,41,42,43)

# combine data files into one matrix

mousedat <- rbind(wt.control[,idx.c],wt.cholate[,idx.c], mut.control[,idx.c],mut.cholate[,idx.c])

# assign column names

names(mousedat) <- c( "strain",
                     "sex",
                     "genotype",
                     "diet",
                     "ast.start", "alt.start", "alp.start",
                     "bili.start", "chol.start",
                     "ast.end", "alt.end","alp.end",
                     "bili.end", "chol.end",
                     "weight.start", "weight.end",
                     "dietdays",
                     "dutchorus",
                     "anes",
                     "dil.start",
                     "dil.end",
                     "hemo.start",
                     "hemo.end",
                     "bile.chol",
                     "bile.phos", "bile.salts",
                     "day0", "day1", "day2","day3", "day4", "day5",
                     "day6", "day7", "day8","day9", "day10", "day11",
                     "day12", "day13","day14", "day15", "day16", "day17",
                     "age", "per.wt.loss",
                     "liverwt", "rel.liverwt")

# set strain, genotype, diet, and sex as factors

mousedat$strain <- as.factor(as.character(mousedat$strain))
mousedat$sex <- as.factor(as.character(mousedat$sex))
mousedat$genotype <- as.factor(as.character(mousedat$genotype))
mousedat$diet <- factor(as.character(mousedat$diet))

# combine the 2 S129 strains
mousedat$strain2 <- mousedat$strain
idx <- grep("S",as.character(mousedat$strain))
mousedat$strain2 <- as.character(mousedat$strain2)
mousedat$strain2[idx] <- "S129"
mousedat$strain2 <- as.factor(mousedat$strain2)

# combine the 2 F1 strains
mousedat$strainC <- mousedat$strain2
idx <- grep("F",as.character(mousedat$strain2))
mousedat$strainC <- as.character(mousedat$strainC)
mousedat$strainC[idx] <- "F1"
mousedat$strainC <- as.factor(mousedat$strainC)

contrasts(mousedat$strainC) <- contr.treatment(3)
 mousedat$diet <- relevel(mousedat$diet,2)
 mousedat$genotype <- relevel(mousedat$genotype,2)

# make fake new data to get estimates effects for each combination of
# four factors

 x1 <- as.factor(rep(c("F","M"),c(12,12)))
 x2 <- rep(c("control","cholate"),c(6,6))
 x2 <- as.factor(c(x2,x2))
 x3 <- rep(c("B6","S129","F1"),c(2,2,2))
 x3 <- as.factor(c(x3,x3,x3,x3))
 x4 <- as.factor(rep(c("WT","mutant"),6))
 
newdata <- data.frame(sex=x1,diet=x2,strainC=x3,genotype=x4)


# Generalized linear mode for stepwise regression

fitbic <- function(formula,data,...)
  {
    git <- glm(as.formula(formula),data=data,...)
    # hit <- step(git,k=log(nrow(git$qr$qr)),trace=0)
    hit <- step(git,k=log(length(git$residuals)),trace=0)
    # print(summary(hit)$call)
    # print(summary(hit)$coefficients,digits=3)
    print(summary(hit),sign=FALSE,digits=2)
    hit
  }

plotpheno <- function(x,y,z,xlab="",ylab="",fac=0,cex=0.7)
  {
    xylim <- range(c(x,fac+y),na.rm=T)
    plot(x,y,xlim=xylim,ylim=xylim-fac,xlab=xlab,ylab=ylab,type="n")
    # open circles
    points(x[z],y[z],pch=21,cex=cex)
    # solid circles
    points(x[!z],y[!z],pch=19,cex=cex)
    lines(xylim,xylim-fac)
  }

form.a <- "strainC*diet*genotype*sex"


idx.days <- ( (mousedat$dietdays>=4) & (mousedat$dietdays<=8) )


# Percent Weight loss perday

# calculate percent weight loss per day
mousedat$per.weight.loss <- (((mousedat$weight.start - mousedat$weight.end)/mousedat$weight.start)*100)/mousedat$dietdays
 
# Fit weight loss data  using fitbit function 

hit3 <- fitbic(per.weight.loss ~ strainC*diet*genotype*sex,data=mousedat )

# Print fitted (predicted) medians for each group of mice (by sex, diet, genotype, and strain)

cbind(newdata,est=round(predict(hit3,newdata,interval="confidence")*1000)/1000)


# Cholesterol

# Several cholesterol data points were marked as \code{<50}.  We treated
# them as missing.

mousedat$chol.start <- as.character(mousedat$chol.start)
idx <- mousedat$chol.start=="<50"

print(paste("Suspect values at start = ", sum(idx,na.rm=T),".",sep=""))

mousedat$start[idx] <- NA
mousedat$chol.start <- as.numeric(mousedat$chol.start)

mousedat$chol.end <- as.character(mousedat$chol.end)
idx <- mousedat$chol.end=="<50"

print(paste("Suspect values at end = ", sum(idx,na.rm=T),".",sep=""))

mousedat$end[idx] <- NA
mousedat$chol.end <- as.numeric(mousedat$chol.end)

form.start <- "log(chol.start) ~"
form.end <- "log(chol.end) ~"
idx.se <- complete.cases(mousedat$chol.start,mousedat$chol.end)

idx <- which(!is.na(log(mousedat$chol.end))&(mousedat$chol.end!=0))

mousedat.compchol<-mousedat[idx,]

# Fit baseline cholesterol data  using fitbit function 
hit0 <- fitbic( as.formula(paste(form.start,form.a)), data=mousedat )

# Print fitted (predicted) medians for each group of mice (by sex, diet, genotype, and #strain)
cbind(newdata,est=round(exp(predict(hit0,newdata,interval="confidence"))*10)/10)

# Fit post-diet cholesterol data  using fitbit function 
hit1 <- fitbic( as.formula(paste(form.end,form.a)), data=mousedat.compchol[idx.days,])

# Print fitted (predicted) medians for each group of mice (by sex, diet, genotype, and #strain)
cbind(newdata,est=round(exp(predict(hit1,newdata,interval="confidence"))*10)/10)


# ALP

mousedat$alp.start <- as.numeric(mousedat$alp.start)

mousedat$alp.end <- as.numeric(mousedat$alp.end)


idx <- which(!is.na(log(mousedat$alp.end))&(mousedat$alp.end!=0))

mousedat.compalp<-mousedat[idx,]


form.start <- "log(alp.start) ~"
form.end <- "log(alp.end) ~"
idx.se <- complete.cases(mousedat$alp.start,mousedat$alp.end)

# Fit baseline ALP data  using fitbit function 
hit0 <- fitbic( as.formula(paste(form.start,form.a)), data=mousedat )

# Print fitted (predicted) medians for each group of mice (by sex, diet, genotype, and #strain)
cbind(newdata,est=round(exp(predict(hit0,newdata,interval="confidence"))*10)/10)

# Fit post-diet ALP data  using fitbit function 
hit1 <- fitbic( as.formula(paste(form.end,form.a)), data=mousedat.compalp[idx.days,] )
# Print fitted (predicted) medians for each group of mice (by sex, diet, genotype, and #strain)
cbind(newdata,est=round(exp(predict(hit1,newdata,interval="confidence"))*10)/10)

# Bilirubin

# For bilirubin we performed a grouped analysis using the proportional
# odds (polr) #function. For baseline we had 2 levels- normal and
# elevated and for post-diet we had 3 #levels- normal, moderately
# elevated, and highly elevated

mousedat$bili.start <- as.character(mousedat$bili.start)
idx <- grep("yellow", mousedat$bili.start)
mousedat$bili.start[idx] <- NA
idx <- grep("<", as.character(mousedat$bili.start))
mousedat$bili.start[idx] <- 0
mousedat$bili.start<- as.numeric(mousedat$bili.start)

mousedat$bili.start[mousedat$bili.start<=0.1] <- 0
mousedat$bili.start[mousedat$bili.start>0.1] <- 1



mousedat$bili.end <- as.character(mousedat$bili.end)
idx <- grep("yellow", mousedat$bili.end)
mousedat$bili.end[idx] <- NA
idx <- grep("<", as.character(mousedat$bili.end))
mousedat$bili.end[idx] <- 0
mousedat$bili.end <- as.numeric(mousedat$bili.end)

mousedat$bili.end[mousedat$bili.end>2.0] <- 3
mousedat$bili.end[mousedat$bili.end>0.1&mousedat$bili.end<=2.0] <- 2
mousedat$bili.end[mousedat$bili.end<=0.1] <- 1

mousedat$bili.end <- as.factor(mousedat$bili.end)



form.start <- "bili.start ~"
form.end <- "bili.end ~"
idx.se <- complete.cases(mousedat$bili.start,mousedat$bili.end)

# Generalized linear model for binomial or 2 level analysis for baseline bilirubin
glm.start<-glm(paste(form.start,form.a), data=mousedat, family=binomial)

# Propotional odds regression for 3 level post-diet bilirubin
pol.end<-polr(paste(form.end,form.a),data=mousedat[idx.days,], weights=Freq)

summary (glm.start)
summary (pol.end)

step(glm.start,k=log(glm.start$df.residual), trace=0)
step(pol.end, k=log(pol.end$df.residual), trace=0)



# Liver Weight relative to final body weight

form.liver<-"log(rel.liverwt)~"

# Fit liver weight at sacrifice using fitbit function 
hit0 <- fitbic( as.formula(paste(form.liver,form.a)), data=mousedat )

#Print fitted (predicted) medians for each group of mice (by sex, diet, genotype, and #strain)
cbind(newdata,est=round(exp(predict(hit0,newdata,interval="confidence"))*10)/10)

# Bile Cholestertol

form.bile <- "log(bile.chol) ~"
hit0 <- fitbic( as.formula(paste(form.bile, form.a)), data=mousedat )
cbind(newdata,est=round(exp(predict(hit0,newdata,interval="confidence"))*10)/10)

# Bile Phospholipids

mousedat$bile.phos<- as.numeric(mousedat$bile.phos)

idx <- which(!is.na(log(mousedat$bile.phos))&(mousedat$bile.phos!=0))
 
mousedat.compbp<-mousedat[idx,]

form.phos <- "log(bile.phos) ~"
hit0 <- fitbic( as.formula(paste(form.phos,form.a)), data=mousedat.compbp )
cbind(newdata,est=round(exp(predict(hit0,newdata,interval="confidence"))*10)/10)

# Bile Salts

mousedat$bile.salts<- as.numeric(mousedat$bile.salts)
idx <- which(!is.na(log(mousedat$bile.salts))&(mousedat$bile.salts!=0))

mousedat.compbs<-mousedat[idx,]
form.salt <- "log(bile.salts) ~"
hit0 <- fitbic( as.formula(paste(form.salt,form.a)), data=mousedat.compbs )
cbind(newdata,est=round(exp(predict(hit0,newdata,interval="confidence"))*10)/10)