\documentclass[11pt]{article}
\usepackage{graphicx}
\usepackage{cite}
\usepackage{hyperref}
\pagestyle{myheadings}
\markright{microfluidicsModels06; Revised: 1 Sept 2010}

\setlength{\topmargin}{0in}
\setlength{\textheight}{8in}
\setlength{\textwidth}{6.5in}
\setlength{\oddsidemargin}{0in}
\setlength{\evensidemargin}{0in}

\def\rcode#1{\texttt{#1}}
\def\fref#1{\textbf{Figure~\ref{#1}}}
\def\tref#1{\textbf{Table~\ref{#1}}}
\def\sref#1{\textbf{Section~\ref{#1}}}

\title{Training a Two-Gene Prognostic Model in CLL}
\author{Kevin R. Coombes}
\date{5 August 2010; Revised 1 September 2010}

\SweaveOpts{prefix.string=Figures/microfluidicsModels06,eps=FALSE}
<<options,echo=FALSE>>=
options(width=92)
options(SweaveHooks = list(fig = function() par(bg='white')),eps=FALSE)
@ 

<<makeFiguresDirectory,echo=FALSE>>=
if (!file.exists("Figures")) {
  dir.create("Figures")
}
@ 

\begin{document}
\maketitle
\tableofcontents

\section{Executive Summary}
\subsection{Introduction}
The main goal of the study is to identify changes in gene expression
in CLL that are associated with clinical outcome (including overall
survival and time-to-treatment).

\subsubsection{Aims/Objectives}
We want to find individual genes and combinations of genes that are
related to prognosis in CLL.  Here ``prognosis'' refers to the ability
to predict time-to-treatment or overall survival.

\subsection{(Statistical) Methods}
Cox proportional hazards models. Kaplan-Meier plots.

\subsection{Results}

The two-gene model using SKI and SLAMF1 is an effective predictor of
time from diagnosis to treatment, or of time from sample to
treatment.  While not significant, it has the correct trend to
potentially predict overall survival after diagnosis.

\subsection{Conclusion}

We should attempt to validate the two-gene model in our validation
dataset.

\section{Loading the Data}

In the previous report (\rcode{microfluidicsModels05.pdf}), we stored
all of the relevant data in a binary R file.  We begin by loading that
data.
<<load.data>>=
load("predictors.rda")
ls()
@ 

\subsection{Preliminaries}

As mentioned previously, we want to perform a variety of time-to-event
(``survival'') analyses, with different starting and ending points.
The basic code for these analyses is in the following libraries.
<<lib>>=
library(survival)
library(BMA)
@ 

\section{Training a two-gene model}

In the previous report, we considered a wide variety of different
models on the training data.  In reviewing the results, we noticed the
following facts:
\begin{enumerate}
\item Many models were able to predict time-to-treatment from either
  diagnosis or sample collection, but were less effective at
  predicting overall survival.  The latter predictions were usually
  better as continuous scores than as binary variables.
\item On reviewing Figure 1 of report \texttt{microfluidicsModels04},
  we noted that the most frequently retained genes to predict time
  from diagnosis to first significant treatment were SKI and SLAMF1.
\item When we used AIC to see which genes were significant in the
  presence of clinical variables that predicted time-to-treatment, the
  same two genes were always retained: SKI and SLAMF1.
\item Of the models we looked at, the smallest one that included SKI
  and SLAMF1 and appeared  to predict overall survival in addition to
  time-to-treatment was \texttt{mod06c}, which included a total of
  four genes: SKI, SLAMF1, CD14, and NT5C2.
\end{enumerate}

In this report, we want to examine the performance (on the training
set) of a model that just uses SKI and SLAMF1 as predictors.

\section{Diagnosis to Treatment}

<<mod02>>=
dset <- data.frame(cardAB.data, cardAB.clinical)
mod02 <- coxph(Surv(TimeDiagnosis2SigTreat, NumericSigTreatment) ~ SKI + SLAMF1, 
               data=dset)
mod02
@ 
The two-gene model is a highly significant continuous predictor of
time from diagnosis to treatment.

Now we compare the two models using a chi-squared test:
<<anova>>=
anova(mod06c, mod02)
@ 
Since this is NOT significant, it suggests that the two-gene model is
as good as the four-gene model.


\subsection{Prognostic scores}

We now convert the predictions into a continuous and a binary prognostic score.
<<scores>>=
x <- datasetAB$mod02 <- predict(mod02)
datasetAB$Cat.mod02 <- factor(ifelse(x > median(x), "HighScore","LowScore"))
rm(x)
@ 
The score remains significant as a continuous or a binary predictor:
<<bmod2>>=
coxph(Surv(TimeDiagnosis2SigTreat, NumericSigTreatment) ~ mod02, datasetAB)
coxph(Surv(TimeDiagnosis2SigTreat, NumericSigTreatment) ~ Cat.mod02, datasetAB)
@ 

<<fig2a.code,eho=FALSE,eval=FALSE>>=
plot(survfit(Surv(TimeDiagnosis2SigTreat, NumericSigTreatment) ~ Cat.mod02, 
             data=datasetAB), col=colorcode, lty=ltype, lwd=2,
     main="Training", 
     xlab="Time (months)", ylab="Fraction Untreated")
legend("topright", levels(datasetAB$Cat.mod02), 
       col=colorcode, lty=ltype, lwd=3)
@ 
<<make.fig2a>>=
colorcode <- "black"
ltype <- c("solid", "dashed")
png(file="SKI-SLAM-paper/mfc-figure2a.png", width=600, height=600,  
    bg="white", pointsize=16)
par(bg="white")
<<fig2a.code>>
dev.off()

pdf(file="SKI-SLAM-paper/mfc-figure2a.pdf", width=6, height=6,
    bg="white", pointsize=12)
par(bg="white")
<<fig2a.code>>
dev.off()
@ 
\begin{figure}
<<fig=TRUE,echo=FALSE>>=
colorcode <- c("red", "blue")
ltype <- "solid"
<<fig2a.code>>
@ 
\caption{Kaplan-Meier plot; training samples; prediction of
  time-to-treatment using two genes.}
\label{dx2rx}
\end{figure}

\section{Overall Survival From Diagnosis}

The prognostic score (from the time-to-treatment analysis) is borderline significant as a continuous
predictor of overall survival:
<<cmod2.os>>=
coxph(Surv(OSAfterDiagnosis, NumericVitalStatus) ~ mod02, datasetAB)
@ 
but it is not significant as a binary predictor:
<<bomd2.os>>=
coxph(Surv(OSAfterDiagnosis, NumericVitalStatus) ~ Cat.mod02, datasetAB)
@ 
However, the trend is in the correct direction, and the quality of the
predictions is not terribly different from the ones we observed
previously when using more genes (\fref{dx2os}).

\begin{figure}
<<fig=TRUE,echo=FALSE>>=
plot(survfit(Surv(OSAfterDiagnosis, NumericVitalStatus) ~ Cat.mod02, 
             data=datasetAB), col=c("red", "blue"), lwd=2,
     main="Training; Overal Survival After Dx (SKI + SLAMF1)", 
     xlab="Time (months)", ylab="Fraction Surviving")
legend("topright", levels(datasetAB$Cat.mod02), col=c("red", "blue"), lwd=3)
@ 
\caption{Kaplan-Meier plot; training samples; prediction of overall
  survival using the score from a time-to-treatment model.}
\label{dx2os}
\end{figure}

\section{Sample Collection to Treatment}

The prognostic score (from the analysis of time from diagnosis to
treatment) is significant as a continuous predictor of time from
sample collection to first significant treatment:
<<cmod2.os>>=
coxph(Surv(TimeSample2SigTreat, NumericSigTreatment) ~ mod02, datasetAB)
@ 
and it remains significant as a binary predictor (\fref{sam2os}):
<<bomd2.os>>=
coxph(Surv(TimeSample2SigTreat, NumericSigTreatment) ~ Cat.mod02, datasetAB)
@ 

\begin{figure}
<<fig=TRUE,echo=FALSE>>=
plot(survfit(Surv(TimeSample2SigTreat, NumericSigTreatment) ~ Cat.mod02, 
             data=datasetAB), col=c("red", "blue"), lwd=2,
     main="Training; Sample to Rx (SKI + SLAMF1)", 
     xlab="Time (months)", ylab="Fraction Untreated")
legend("topright", levels(datasetAB$Cat.mod02), col=c("red", "blue"), lwd=3)
@ 
\caption{Kaplan-Meier plot; training samples; prediction of overall
  survival using the score from a time-to-treatment model.}
\label{sam2os}
\end{figure}

\section{Appendix}

<<save.me>>=
modlist <- c(modlist, "mod02")
cutoffs <- c(cutoffs, median(datasetAB$mod02))
save.image(file="withmod02.rda")
@ 

This analysis was run in the following directory:
<<getwd>>=
getwd()
@

<<lib2,echo=F>>=
sessionInfo()
@ 

\end{document}