%\VignetteIndexEntry{SVAPLSseq tutorial}
% \VignetteKeywords{Gene expression data, Normalization, RNA-seq, batch effects}
% \VignettePackage{SVAPLSseq}
\documentclass[12pt]{article}
<<echo=FALSE>>=
options(width=65)
@ 
<<style-Sweave, eval=TRUE, echo=FALSE, results=tex>>=
BiocStyle::latex()
@

\SweaveOpts{eps=FALSE,echo=TRUE}

\begin{document}
\SweaveOpts{concordance=TRUE}
\title{SVAPLSseq: An R package to correct for hidden sources of variability in differential gene expression studies based on RNAseq data}
\author{Sutirtha Chakraborty$^1$* \\
$^1$National Institute of Biomedical Genomics 
*email: \texttt{sc4@nibmg.ac.in}}
\date{Modified: June 25, 2016  Compiled: \today}
\maketitle

\tableofcontents
\section{Overview}

The R package \Rpackage{SVAPLSseq} contains functions that are intended for the extraction and correction of different types of hidden 
biological and technical variables that could potentially generate latent heterogeneity in RNAseq data on gene expression. The complexity 
of the sequencing workflow creates a number of technical artefacts along with the inherent biological variability stemming from the 
unknown gene and sample profiles. The package aims to provide the users with a flexible and generalized framework to identify these 
hidden effects and adjust for them in order to re-estimate the primary signals of group-specific differential gene expression with 
higher power and accuracy. The underlying method operates by implementing a non-linear partial least squares regression algorithm on 
two multivariate random matrices constructed from the data. To that end two methodological variants are provided in this package: 
(1) Unsupervised SVAPLSseq and (2) Supervised SVAPLSseq. Both these variants yield a set of surrogate variables that are then tested 
for statistical significance in order to detect the important signatures of latent variability in the data. The package also provides 
an added functionality in terms of incorporating these extracted signatures in a linear regression framework and estimating the group-
specific differential expression effects. For this purpose two different options are provided: (a) Wald test that uses the R packages 
``edgeR'' and ``limma'' and (b) Likelihood ratio test. 

This document provides a tutorial to use the package for:
 
\begin{itemize}
\item Formatting the data for use in the package.\\
\item Extracting the signatures of the hidden effects in the data.\\
\item Using the estimated hidden effect signatures to detect the truly differentially expressed genes. 
\end{itemize}

\section{Formatting the data for use in the package}

The starting step for using the package is to set up the RNAseq expression data in an appropriate format. The input data should be in 
the form of either a count matrix object or a 'SummarizedExperiment' or a 'DGEList' object. The object will contain a feature matrix 
that will list the features (genes/transcripts) along the rows and samples along the columns. This matrix will contain the read count 
values for the features corresponding to the different samples. In addition, a separate factor variable should be designed that will 
keep track of the group each sample belongs (e.g. ``treated'' and ``untreated'', ``Normal'' and ``Cancer''). This variable will enable 
the estimation of the primary signal for group-specific differential expression of the features.

<<input>>=
library(SummarizedExperiment)
library(SVAPLSseq)
library(edgeR)

data(sim.dat)
dat = SummarizedExperiment(assays = SimpleList(counts = sim.dat))
dat = DGEList(counts = sim.dat)
sim.dat[1:6, c(1:3, 11:13)]
@

\section{Extracting the signatures of the hidden effects in the data}

The package contains a function \Rfunction{svplsSurr} that extracts the signatures of latent variability (surrogate variables) in the 
data by using a multivariate non-linear partial least squares (NPLS) algorithm (Boulesteix and Strimmer 2007). The function takes the 
original read count matrix of feature expression values along with a factor variable indicating the group of each sample as input. 
Moreover, it allows the user to specify a certain number of surrogate variables (\texttt{max.surrs}) that will be extracted from the 
data. These variables are further tested for statistical significance to generate an optimal set of significant surrogate variables 
capturing the latent variation in the data. The function returns a matrix with these variables along the columns and a vector 
containing the proportions of the total variance in the data space that are explained by them.\\

\noindent
The function provides the user with two methodological variants: (1) The Unsupervised SVAPLSseq and (2) The Supervised SVAPLSseq. 
Details on these two variants and their usage on an RNAseq gene expression data are provided below:

\subsection{The Unsupervised SVAPLSseq}

This version of the method regresses the primary signal corrected residual matrix on the original gene expression data matrix via NPLS. 
The estimated scores in the data space are considered as the surrogate variables that are further tested for statistical significance. 
Setting the \texttt{controls} argument of the function to \texttt{NULL} starts this version.

<<input>>=
data(sim.dat)
group = as.factor(c(rep(1, 10), rep(-1, 10)))
sim.dat.se = SummarizedExperiment(assays = SimpleList(counts = sim.dat))
sim.dat.dg = DGEList(counts = sim.dat)

sv <- svplsSurr(dat = sim.dat, group = group, max.surrs = 3, controls = NULL)
sv <- svplsSurr(dat = sim.dat.se, group = group, max.surrs = 3, controls = NULL)
sv <- svplsSurr(dat = sim.dat.dg, group = group, max.surrs = 3, controls = NULL)
print(sv)
surr(sv)             
prop.vars(sv)        
@

\subsection{The Supervised SVAPLSseq}

In this variant a submatrix of the original gene expression data is first created corresponding to a set of available control probes 
that do not have any differential expression between the two groups. Hence, this submatrix is only expected to contain the signatures 
of the hidden effects in the data. This matrix is then regressed on the original data to extract the surrogate variables for the 
underlying latent variation. This variant is called by setting the \texttt{controls} argument of the function to the collection of the 
control probes.

<<input>>=
data(sim.dat)
controls = c(1:nrow(sim.dat)) > 400
group = as.factor(c(rep(1, 10), rep(-1, 10)))
sim.dat.se = SummarizedExperiment(assays = SimpleList(counts = sim.dat))
sim.dat.dg = DGEList(counts = sim.dat)

sv <- svplsSurr(dat = sim.dat, group = group, max.surrs = 3, controls = controls)
sv <- svplsSurr(dat = sim.dat.se, group = group, max.surrs = 3, controls = controls)
sv <- svplsSurr(dat = sim.dat.dg, group = group, max.surrs = 3, controls = controls)
print(sv)
surr(sv)           
prop.vars(sv)   
@

\section{Using the estimated hidden effect signatures to detect the truly differentially expressed genes}

The package contains another function \Rfunction{svplsTest} that incorporates the significant surrogate variables estimated by the 
function \Rfunction{svplsSurr} inside a regression framework in order to test for the genes that are truly differentially expressed 
between the two groups. The function provides the user with two testing options: (1) Wald test based on the regression coefficients 
of the primary signal effects (group effects) after incorporating the surrogate variables in a linear model and (2) Likelihood ratio 
test (LRT) comparing two different regression models: one containing  primary signal effects as well as the surrogate variables and 
the other including only the surrogate variables. A list is returned as the output that contains the genes detected to be differentially
expressed between the two groups (\texttt{sig.genes}, the uncorrected pvalues from the test (\texttt{pvs.unadj}) and the corresponding 
FDR adjusted pvalues (\texttt{pvs.adj}).

<<input>>=
data(sim.dat)
group = as.factor(c(rep(1, 10), rep(-1, 10)))
sv = svplsSurr(dat = sim.dat, group = group)
surr = surr(sv)

sim.dat.se = SummarizedExperiment(assays = SimpleList(counts = sim.dat))
sim.dat.dg = DGEList(counts = sim.dat)

fit = svplsTest(dat = sim.dat, group = group, surr = surr, test = "Wald")
fit = svplsTest(dat = sim.dat.se, group = group, surr = surr, test = "Wald")
fit = svplsTest(dat = sim.dat.dg, group = group, surr = surr, test = "Wald")

sig.genes(fit) 
pvs.unadj(fit)   
pvs.adj(fit)     
@  

\begin{thebibliography}{}
\bibitem{boule1} 
Boulesteix, A. L. and Strimmer, K. (2007) Partial least squares: a versatile tool for the analysis of high-dimensional genomic data. 
{\it Briefings in Bioinformatics} {\bf 8(1),} 32--44.
\end{thebibliography}


\end{document}