% \VignetteEngine{knitr::knitr}
% \VignetteIndexEntry{Annotation -- slides}

\documentclass{beamer}
\usepackage{BioconductorSlides}
\title{Gene Set Enrichment -- Introduction}
\author{Martin Morgan 
  (\href{mailto:martin.morgan@roswellpark.org}{martin.morgan@roswellpark.org}) \\
  Roswell Park Cancer Institute \\
  Buffalo, NY, USA
}
\date{3 July, 2019}

\begin{document}

\maketitle

\begin{frame}{Objective}
  Is expression of genes in a gene set associated with experimental
  condition?
  \begin{itemize}
  \item E.g., Are there unusually many up-regulated genes in the gene
    set?
  \end{itemize}
  Many methods, a review is Kharti et al., 2012.
  \begin{itemize}
  \item Over-representation analysis (ORA) -- are differentially
    expressed (DE) genes in the set more common than expected?
  \item Functional class scoring (FCS) -- summarize statistic of DE of
    genes in a set, and compare to null
  \item Issues with sequence data?
  \item Issues with single-cell data?
  \end{itemize}
\end{frame}

\begin{frame}{What is a gene set?}
  \textbf{Any} \emph{a priori} classification of `genes' into
  biologically relevant groups
  \begin{itemize}
  \item Members of same biochemical pathway
  \item Proteins expressed in identical cellular compartments
  \item Co-expressed under certain conditions
  \item Targets of the same regulatory elements
  \item On the same cytogenic band
  \item \ldots
  \end{itemize}
  Sets do not need to be\ldots
  \begin{itemize}
  \item \emph{exhaustive}
  \item \emph{disjoint}
  \end{itemize}
\end{frame}

\begin{frame}{Collections of gene sets}
  Gene Ontology (\href{http://geneontology.org}{GO}) Annotation (GOA)
  \begin{itemize}
  \item CC Cellular Components
  \item BP Biological Processes
  \item MF Molecular Function
  \end{itemize}
  Pathways
  \begin{itemize}
  \item \href{http://www.broadinstitute.org/gsea/msigdb/}{MSigDb}
  \item \href{http://genome.jp/kegg}{KEGG}
  \item \href{http://reactome.org}{reactome}
  \item \href{http://pantherdb.org}{PantherDB}
  \item \ldots
  \end{itemize}
\end{frame}

\begin{frame}{Collections of gene sets}
  E.g., \href{http://www.broadinstitute.org/gsea/msigdb/}{MSigDb}
  \begin{itemize}
  \item c1 Positional gene sets -- chromosome \& cytogenic band
  \item c2 Curated Gene Sets from online pathway databases,
    publications in PubMed, and knowledge of domain experts.
  \item c3 motif gene sets based on conserved cis-regulatory motifs
    from a comparative analysis of the human, mouse, rat, and dog
    genomes.
  \item c4 computational gene sets defined by mining large collections
    of cancer-oriented microarray data.
  \item c5 GO gene sets consist of genes annotated by the same GO
    terms.
  \item c6 oncogenic signatures defined directly from microarray gene
    expression data from cancer gene perturbations.
  \item c7 immunologic signatures defined directly from microarray
    gene expression data from immunologic studies.
  \end{itemize}
\end{frame}

\begin{frame}{Work flow}
  \begin{enumerate}
  \item Experimental design
  \item Sequencing, quality assessment, alignment
  \item Differential expression
  \end{enumerate}
  and then\ldots
  \begin{enumerate}
    \setcounter{enumi}{3}
  \item Perform gene set enrichment analysis
  \item Adjust for multiple comparisons
  \end{enumerate}
\end{frame}

\begin{frame}{Approach 1: hypergeometric tests}
  \begin{columns}
    \column{.5\textwidth}
    \begin{enumerate}
    \item Classify each gene as `differentially expressed' DE or not,
      e.g., based on $p < 0.05$
    \item Are DE genes in the set more common than DE genes not in the
      set?
    \item Fisher hypergeometric test. \Biocpkg{GOstats};
      \Biocpkg{limma}\Rfunction{::goana()} \ldots
    \end{enumerate}
    \begin{itemize}
    \item Conditional hypergeometric to accommodate GO DAG,
      \Biocpkg{GOstats}
    \item But: artificial division into two groups (DE vs.\ not DE)
    \end{itemize}
    \column{.5\textwidth}
    \begin{tabular}{lcc}
             & \multicolumn{2}{c}{In gene set?}\\
             & Yes   & No \\ \hline\noalign{\smallskip}
      DE     & $k$   & $K$ \\
      Not DE & $n-k$ & $N - K$ \\ \hline
    \end{tabular}
    \par\bigskip \Rfunction{fisher.test()}
  \end{columns}
\end{frame}

\begin{frame}{Approach 2: enrichment score}
  \begin{columns}
    \column{.5\textwidth}
    Mootha et al., 2003; modified Subramanian et al., 2005.
    \begin{enumerate}
    \item Sort genes by log fold change
    \item Calculate running sum: incremented when gene in set,
      decremented when not.
    \item Maximum of the running sum is enrichment score ES; large ES
      means that genes in set are toward top of list.
    \item Permuting subject labels for signficance
    \end{enumerate}
    \column{.5\textwidth}
    \includegraphics[width=\textwidth]{our_figures/subramanian-F1-part.jpg}
    \par {\small Subramanian et al., 2005, fig 1.}
  \end{columns}
\end{frame}

\begin{frame}{Approach 3: category $t$-test}
  \begin{columns}
    \column{.5\textwidth}
    E.g., Jiang \& Gentleman, 2007; \Biocpkg{Category}
    \begin{enumerate}
    \item Summarize $t$ (or other) statistic across genes in each set
    \item Test for significance by permuting the subject labels
    \item Much more straight-forward to implement
    \end{enumerate}
    \column{.5\textwidth}
    \includegraphics[width=\textwidth]{our_figures/Category-ribosome.png}
    \par {\small Expression in NEG vs BCR/ABL samples for genes in the
      `ribosome' KEGG pathway; \Biocpkg{Category} vignette.}
  \end{columns}
\end{frame}

\begin{frame}{Competitive versus self-contained null hypothesis}
  Goemann \& B\"uhlmann, 2007
  \begin{itemize}
  \item Competitive null: The genes in the gene set do not have
    stronger association with the subject condition than other
    genes. Distinguishing more from less important sets. (Approach 1,
    2)
  \item Self-contained null: The genes in the gene set do not have any
    association with the subject condition. Assessing individual
    sets. (Approach 3)
  \item Probably, self-contained null is closer to actual question of
    interest
  \item Permuting subjects (rather than genes) is appropriate
  \end{itemize}
\end{frame}

\begin{frame}{Approach 4: linear models}
  E.g., Hummel et al., 2008, \Biocpkg{GlobalAncova}
  \begin{itemize}
  \item Colorectal tumors have good (`stage II') or bad (`stage III')
    prognosis. Do genes in the p53 pathway (\emph{just one gene set!})
    show different activity at the two stages?
  \item Linear model incorporates covariates -- sex of patient,
    location of tumor
  \end{itemize}
  \Biocpkg{limma}
  \begin{itemize}
  \item Majewski et al., 2010 \Rfunction{romer()} and Wu \& Smythe 2012
    \Rfunction{camera()} for enrichment (competitive null) linear models
  \item Wu et al., 2010: \Rfunction{roast()}, \Rfunction{mroast()},
    (and \Rfunction{fry()} -- efficient) for self-contained null
    linear models
  \end{itemize}
\end{frame}

\begin{frame}{Approach 5: issues with sequence data?}
  \begin{columns}
    \column{.5\textwidth}
    \begin{itemize}
    \item All else being equal, long genes receive more reads than
      short genes
    \item Per-gene $P$ values proportional to gene size
    \end{itemize}
    E.g., Young et al., 2010, \Biocpkg{goseq}
    \begin{itemize}
    \item Hypergeometric, weighted by gene size
    \item Substantial differences
    \item Better: read depth??
    \end{itemize}
    \column{.5\textwidth}
    \includegraphics[width=\textwidth]{our_figures/young-et-al-gb-2010-11-2-r14-2-cropped.png}
    \par {\small DE genes vs.\ transcript length. Points: bins of 300
      genes. Line: fitted probability weighting function.}
  \end{columns}
\end{frame}

\begin{frame}{Approach 6: \emph{de novo} discovery}
  \begin{itemize}
  \item So far: analogous to supervised machine learning, where
    pathways are known in advance
  \item What about unsupervised discovery?
  \end{itemize}
  Example: Langfelder \& Hovarth, WGCNA
  \begin{itemize}
  \item Weighted correlation network analysis
  \item Described in Langfelder \& Horvath, 2008
  \end{itemize}
\end{frame}

\begin{frame}{Issues with single-cell data?}
  \begin{itemize}
  \item Often, projections into reduced dimensions.
  \item Not genes \emph{per se}, but weightings.
  \item An open issue, with opportunities for new methods!
  \end{itemize}
\end{frame}

\begin{frame}{Representing gene sets in R}
  \begin{itemize}
  \item Named \Rcode{list()}, where names of the list are sets, and
    each element of the list is a vector of genes in the set.
  \item \Rcode{data.frame()} of set name / gene name pairs
  \item \Biocpkg{GSEABase} -- input from standard file formats,
    representation as formal classes.
  \end{itemize}
\end{frame}

\begin{frame}{Benchmarks}
  A recent
  \href{https://twitter.com/LeviWaldron1/status/1142092301403115521}{tweet}
  from Levi Waldron provides a nice summary.
  \begin{itemize}
  \item \Biocpkg{GSEABenchmarkR} for running benchmarks
  \item Self-contained tests often call random gene sets significant
  \item Hypergeometric test performs relatively well!
  \end{itemize}
\end{frame}

\begin{frame}{Conclusions}
  Gene set enrichment classifications
  \begin{itemize}
  \item Kharti et al: Over-representation analysis; functional class
    scoring; pathway topology
  \item Goemann \& B\"uhlmann: Competitive vs.\ self-contained null
  \end{itemize}
  \medskip\par Selected \Bioconductor{} packages (see
  \href{http://bioconductor.org/packages/release/BiocViews.html\#___GeneSetEnrichment}{biocViews})
  \medskip\par
  \begin{tabular}{ll}
    Approach & Packages \\
    \hline\noalign{\smallskip}
    Hypergeometric & \Biocpkg{GOstats}, \Biocpkg{topGO},
                     \Biocpkg{limma}\Rfunction{::goana()} \\
    Enrichment & \Biocpkg{limma}\Rfunction{::romer()} \\
    Category $t$-test & \Biocpkg{Category} \\
    Linear model & \Biocpkg{GlobalAncova}, \Biocpkg{GSEAlm}, 
                   \Biocpkg{limma}\Rfunction{::fry()} \\
    Pathway topology & \Biocpkg{SPIA} \\
    Sequence-specific & \Biocpkg{goseq} \\
    Visualization & \Biocpkg{pathview} \\
    \hline
  \end{tabular}
\end{frame}

\begin{frame}{References}
  \begin{itemize}
  \item Khatri et al., 2012, PLoS Comp Biol 8.2: e1002375.
  \item Subramanian et al., 2005, PNAS 102.43: 15545-15550.
  \item Jiang \& Gentleman, 2007, Bioinformatics Feb 1;23(3):306-13.
  \item Goeman \& B\"uhlmann, 2007, Bioinformatics 23.8: 980-987.
  \item Hummel et al., 2008, Bioinformatics 24.1: 78-85.
  \item Wu \& Smyth 2012, Nucleic Acids Research 40, e133.
  \item Wu et al., 2010 Bioinformatics 26, 2176-2182.
  \item Majewski et al., 2010, Blood, published online 5 May 2010.
  \item Tarca et al., 2009, Bioinformatics 25.1: 75-82.
  \item Young et al., 2010, Genome Biology 11:R14.
  \end{itemize}
  Partly based on a presentation by Simon Anders, CSAMA
  2010\footnote{\url{http://marray.economia.unimi.it/2009/material/lectures/L8_Ge
      ne_Set_Testing.pdf}}.
\end{frame}

\begin{frame}{Acknowledgments}
  \begin{itemize}
  \item Core: Qian Liu, Kayla Morrel, Herv\'e Pag\`es, Lori Shepherd,
    Marcel Ramos, Nitesh Turaga, Daniel van Twisk, Jeifie Wang.
  \item The research reported in this presentation was supported by
    the National Cancer Institute and the National Human Genome
    Research Institute of the National Institutes of Health under
    Award numbers U24CA180996 and U41HG004059. The content is
    solely the responsibility of the authors and does not necessarily
    represent the official views of the National Institutes of Health
    or the National Science Foundation.
  \end{itemize}
  \medskip\par
  \url{https://bioconductor.org},
  \url{https://support.bioconductor.org}
\end{frame}

\end{document}