%\VignetteIndexEntry{transcription factor binding site (TFBS) analysis with the "TFBSTools" package} %\VignettePackage{TFBSTools} %\VignetteEngine{knitr::knitr} \documentclass[a4paper,11pt]{article} <>= library(knitr) ## set global chunk options opts_chunk$set(fig.path='figure/manual-', cache.path='cache/manual-', fig.align='center', fig.show='hold', par=TRUE) ## I use = but I can replace it with <-; set code/output width to be 68 options(replace.assign=TRUE, width=68, digits=4) ## tune details of base graphics (http://yihui.name/knitr/hooks) knit_hooks$set(par=function(before, options, envir){ if (before && options$fig.show!='none') par(mar=c(4,4,.1,.1),cex.lab=.95,cex.axis=.9,mgp=c(2,.7,0),tcl=-.3) }) @ <>= BiocStyle::latex() @ <>= library("TFBSTools") @ \usepackage{amssymb,amsmath} \bioctitle{The \Biocpkg{TFBSTools} package overview} \author{Ge Tan \footnote{ge.tan09@imperial.ac.uk}} \date{Edited: Jan 2015; Compiled: \today} \begin{document} \maketitle \tableofcontents \section{Introduction} Eukaryotic regulatory regions are characterized based a set of discovered transcription factor binding sites (TFBSs), which can be represented as sequence patterns with various degree of degeneracy. This \Biocpkg{TFBSTools} package is designed to be a compuational framework for TFBSs analysis. Based on the famous perl module TFBS \cite{lenhard_tfbs:_2002}, we extended the class definitions and enhanced implementations in an interactive environment. So far this package contains a set of integrated \R{} \Rcode{S4} style classes, tools , JASPAR database interface functions. Most approaches can be described in three sequential phases. First, a pattern is generated for a set of target sequences known to be bound by a specific transcription factor. Second, a set of DNA sequences are analyzed to determine the locations of sequences consistent with the described binding pattern. Finally, in advanced cases, predictive statistical models of regulatory regions are constructed based on mutiple occurrences of the detected patterns. \Biocpkg{TFBSTools} aims to support all these functionalities in the environment \R{}, except the external motif finding software, such as \software{MEME} \cite{bailey_fitting_1994}. \section{\Rcode{S4} classes in TFBSTools} The package is built around a number of \Rcode{S4} class of which the \Rclass{XMatrix}, \Rclass{SiteSet} classes are the most important. The section will briefly explain most of them defined in \Biocpkg{TFBSTools}. \subsection{XMatrix and its subclasses} \Rclass{XMatrix} is a virtual class, which means no concrete objects can be created directly from it. The subclass \Rclass{PFMatrix} is designed to store all the relevant information for one raw position frequency matrix (PFM). This object is compatible with one record from JASPAR database. \Rclass{PWMatrix} is used to store a position weight matrix (PWM). Compared with \Rclass{PFMatrix}, it has one extra slot \Rcode{pseudocounts}. \Rclass{ICMatrix} is used to store a information content matrix (ICM). Compared with \Rclass{PWMatrix}, it has one extra slot \Rcode{schneider}. The following examples demonstrate the creation of \Rclass{PFMatrix}, the conversions between these matrices and some assocated methods defined for these classes. <>= ## PFMatrix construction; Not all of the slots need to be initialised. pfm = PFMatrix(ID="MA0004.1", name="Arnt", matrixClass="Zipper-Type", strand="+", bg=c(A=0.25, C=0.25, G=0.25, T=0.25), tags=list(family="Helix-Loop-Helix", species="10090", tax_group="vertebrates",medline="7592839", type="SELEX", ACC="P53762", pazar_tf_id="TF0000003", TFBSshape_ID="11", TFencyclopedia_ID="580"), profileMatrix=matrix(c(4L, 19L, 0L, 0L, 0L, 0L, 16L, 0L, 20L, 0L, 0L, 0L, 0L, 1L, 0L, 20L, 0L, 20L, 0L, 0L, 0L, 0L, 20L, 0L), byrow=TRUE, nrow=4, dimnames=list(c("A", "C", "G", "T"))) ) pfm ## coerced to matrix as.matrix(pfm) ## access the slots of pfm ID(pfm) name(pfm) Matrix(pfm) ncol(pfm) length(pfm) ## convert a FPM to PWM, ICM pwm = toPWM(pfm, type="log2probratio", pseudocounts=0.8, bg=c(A=0.25, C=0.25, G=0.25, T=0.25)) icm = toICM(pfm, pseudocounts=sqrt(rowSums(pfm)[1]), schneider=FALSE, bg=c(A=0.25, C=0.25, G=0.25, T=0.25)) ## get the reverse complment matrix with all the same information except the strand. pwmRevComp <- reverseComplement(pwm) @ \subsection{XMatrixList and its subclasses} \Rclass{XMatrixList} is used to store a set of \Rclass{XMatrix} objects. Basically it is a SimpleList for easy manipulation the whole set of \Rclass{XMatrix}. The concrete objects can be \Rclass{PFMatrix}, \Rclass{PWMatrix} and \Rclass{ICMatrix}. <>= pfm2 = pfm pfmList = PFMatrixList(pfm1=pfm, pfm2=pfm2, use.names=TRUE) pfmList names(pfmList) @ \subsection{SiteSet, SiteSetList, SitePairSet and SitePairSetList} The \Rclass{SiteSet} class is a container for storing a set of putative transcription factor binding sites on a nucleotide sequence (start, end, strand, score, pattern as a \Rclass{PWMatrix}, etc.) from scaning a nucleotide sequence with the corresponding \Rclass{PWMatrix}. Similarly, \Rclass{SiteSetList} stores a set of \Rclass{SiteSet} objects. For holding the results returned from a pairwise alignment scaaning, \Rclass{SitePairSet} and \Rclass{SitePairSetList} are provided. More detailed examples of using these classes will be given in Section \ref{sec:Scanning}; \subsection{MotifSet} This \Rclass{MotifSet} class is used to store the generated motifs from \textit{de novo} motif discovery software, such as \software{MEME} \cite{bailey_fitting_1994}. \section{Database interfaces for JASPAR2014 database} This section will demonstrate how to operate on the JASPAR 2014 database. JASPAR is a collection of transcription factor DNA-binding preferences, modeled as matrices. These can be converted into PWMs, used for scanning genomic sequences. JASPAR is the only database with this scope where the data can be used with no restrictions (open-source). A \Bioconductor{} experiment data package \Biocexptpkg{JASPAR2014} is provided with the leatest release of JASPAR. \subsection{Search JASPAR2014 database} This search function fetches matrix data for all matrices in the database matching criteria defined by the named arguments and returns a PFMatrixList object. For more search criterias, please see the help page for \Rfunction{getMatrixSet}. <>= library(JASPAR2014) opts = list() opts[["species"]] = 9606 opts[["name"]] = "RUNX1" #opts[["class"]] = "Ig-fold" opts[["type"]] = "SELEX" opts[["all_versions"]] = TRUE PFMatrixList = getMatrixSet(JASPAR2014, opts) PFMatrixList opts2 = list() opts2[["type"]] = "SELEX" PFMatrixList2 = getMatrixSet(JASPAR2014, opts2) PFMatrixList2 @ \subsection{Store, delete and initialize JASPAR2014 database} We also provide some functions to initialize an empty JASPAR2014 style database, store new \Rclass{PFMatrix} or \Rclass{PFMatrixList} into it, or delete some records based on ID. The backend of the database is \software{SQLite}. <>= db = "myMatrixDb.sqlite" initializeJASPARDB(db) storeMatrix(db, pfm) deleteMatrixHavingID(db, "MA0003") @ \section{PFM, PWM and ICM methods} This section will give an introduction of matrix operations, including conversion from PFM to PWM and ICM, profile matrices comparison, dynamic random profile generation. \subsection{PFM to PWM} The method \Rfunction{toPWM} can convert PFM to PWM \cite{Wasserman:2004ec}. Optional parameters include \Rcode{type}, \Rcode{pseudocounts}, \Rcode{bg}. The implementation in this package is a bit different from that in \Biocpkg{Biostrings}. First of all, \Rfunction{toPWM} allows the input matrix to have different column sums, which means the count matrix can have an unequal number of sequences contributing to each column. This scenario is rare, but exists in JASPAR SELEX data. Second, we can specify customized \Rcode{pseudocounts}. \Rcode{pseudocounts} is necessary for correcting the small number of counts or eliminating the zero values before log transformation. In TFBS perl module, the square root of the number of sequences contributing to each column. However, it has been shown to too harsh \cite{nishida_pseudocounts_2009}. Hence, a default value of 0.8 is used. Of course, it can be changed to other customized value or even different values for each column. <>= pwm = toPWM(pfm, pseudocounts=0.8) pwm @ \subsection{PFM to ICM} The method \Rcode{toICM} can convert PFM to ICM \cite{schneider_information_1986}. Besides the similar \Rcode{pseudocounts}, \Rcode{bg}, you can also choose to do the \Rcode{schneider} correction. The information content matrix has a column sum between 0 (no base preference) and 2 (only 1 base used). Usually this information is used to plot sequence log. How a PFM is converted to ICM: we have the PFM matrix $x$, base backrgound frequency $bg$, $pseudocounts$ for correction. \begin{align} Z[j] &= \sum_{i=1}^{4} x[i,j] \\ p[i,j] &= {(x[i,j] + bg[i] \times pseudocounts[j]) \over (Z[j] + \sum_{i} bg[i] \times pseudocounts[j]} \\ D[j] &= \log_2{4} + \sum_{i=1}^{4} p[i,j]*\log{p[i,j]} \\ ICM[i,j] &= p[i,j] \times D[j] \end{align} <>= icm = toICM(pfm, pseudocounts=0.8, schneider=TRUE) icm @ To plot the sequence logo, we use the package \Biocpkg{seqlogo}. <>= seqLogo(icm) @ \subsection{Align PFM to a custom matrix or IUPAC string} In some cases, it is beneficial to assess similarity of existing profile matrices, such as JASPAR, to a newly discovered matrix (as with using BLAST for sequence data comparison when using Genbank). \Biocpkg{TFBSTools} provides tools for comparing pairs of PFMs, or a PFM with IUPAC string, using a modified Needleman-Wunsch algorithm \cite{sandelin_integrated_2003}. <>= ## one to one comparison data(MA0003.2) data(MA0004.1) pfmSubject = MA0003.2 pfmQuery = MA0004.1 PFMSimilarity(pfmSubject, pfmQuery) ## one to several comparsion PFMSimilarity(pfmList, pfmQuery) ## align IUPAC string IUPACString = "ACGTMRWSYKVHDBN" PFMSimilarity(pfmList, IUPACString) @ \subsection{PWM similarity} To measure the similarity of two PWM matrix in three measurements: ``normalised Euclidean distance'', ``Pearson correlation'' and ``Kullback Leibler divergence'' \cite{linhart_transcription_2008}. Given two PWMs, $P^1$ and $P^2$, where l is the length. $P_{i,b}$ is the values in column i with base b. The normalised Euclidean distance is computed in \begin{equation} D(a,b) = {1 \over {\sqrt{2}l}} \cdot \sum_{i=1}^{l} \sqrt{\sum_{b \in {\{A,C,G,T\}}} (P_{i,b}^1-P_{i,b}^2)^2} \end{equation} This distance is between 0 (perfect identity) and 1 (complete dis-similarity). The pearson correlation coefficient is computed in \begin{equation} r(P^1, P^2) = {1 \over l} \cdot \sum_{i=1}^l {\sum_{b \in \{A,C,G,T\}} (P_{i,b}^1 - 0.25)(P_{i,b}^2-0.25) \over \sqrt{\sum_{b \in \{A,C,G,T\}} (P_{i,b}^1 - 0.25)^2 \cdot \sum_{b \in \{A,C,G,T\}} (P_{i,b}^2 - 0.25)^2}}. \end{equation} The Kullback-Leibler divergence is computed in \begin{equation} KL(P^1, P^2) = {1 \over {2l}} \cdot \sum_{i=1}^l \sum_{b \in \{A,C,G,T\}} (P_{i,b}^1\log{ P_{i,b}^1 \over P_{i,b}^2}+ P_{i,b}^2\log{P_{i,b}^2 \over {P_{i,b}^1}}). \end{equation} <>= data(MA0003.2) data(MA0004.1) pwm1 = toPWM(MA0003.2, type="prob") pwm2 = toPWM(MA0004.1, type="prob") PWMSimilarity(pwm1, pwm2, method="Euclidean") PWMSimilarity(pwm1, pwm2, method="Pearson") PWMSimilarity(pwm1, pwm2, method="KL") @ \subsection{Dynamic random profile generation} In this section, we will demonstrate the capability of random profile matrices generation with matrix permutation and probabilitis sampling. In many computational/simulation studies, it is particularly desired to have a set of random matrices. Some cases includes the estimation of distance between putative TFBS and transcription start site, the evaluation of comparison between matrices \cite{bryne_jaspar_2008}. These random matrices are expected to have same statistical properties with the selcted profiles, such as nucleotide content or information content. The permutation method is relatively easy. It simply shuffles the columns either constrainted in each matrix, or columns almong all selected matrices. The probabilistic sampling is more complicated and can be done in two steps: \begin{enumerate} \item A Dirichlet multinomial mixture model is trained on all available matrices in JASPAR. \item Random columns are sampled from the posterior distribution of the trained Dirichlet model based on selected profiles. \end{enumerate} <>= ## Matrice permutation permuteMatrix(pfmQuery) permuteMatrix(pfmList, type="intra") permuteMatrix(pfmList, type="inter") @ <>= ## Dirichlet model training data(MA0003.2) data(MA0004.1) pfmList <- PFMatrixList(pfm1=MA0003.2, pfm2=MA0004.1, use.names=TRUE) dmmParameters <- dmmEM(pfmList, K=6, alg="C") ## Matrice sampling from trained Dirichlet model pwmSampled <- rPWMDmm(MA0003.2, dmmParameters$alpha0, dmmParameters$pmix, N=1, W=6) @ \section{Scan sequence and alignments with PWM pattern} \label{sec:Scanning} \subsection{searchSeq} \Rfunction{searchSeq} scans a nucleotide sequence with the pattern represented in the PWM. The strand argument controls which strand of the sequence will be searched. When it is ``*'', both strands will be scanned. A \Rclass{siteset} object will be returned which can be exported into GFF3 or GFF2 format. Empirical p-values for the match scores can be calculated by an exact method from \CRANpkg{TFMPvalue} or the distribution of sampled scores. <>= library(Biostrings) data(MA0003.2) data(MA0004.1) pwmList = PWMatrixList(MA0003.2=toPWM(MA0003.2), MA0004.1=toPWM(MA0004.1), use.names=TRUE) subject = DNAString("GAATTCTCTCTTGTTGTAGTCTCTTGACAAAATG") siteset = searchSeq(pwm, subject, seqname="seq1", min.score="60%", strand="*") sitesetList = searchSeq(pwmList, subject, seqname="seq1", min.score="60%", strand="*") ## generate gff2 or gff3 style output head(writeGFF3(siteset)) head(writeGFF3(sitesetList)) head(writeGFF2(siteset)) ## get the relative scores relScore(siteset) relScore(sitesetList) ## calculate the empirical p-values of the scores pvalues(siteset, type="TFMPvalue") pvalues(siteset, type="sampling") @ \subsection{searchAln} \Rfunction{searchAln} scans a pairwise alignment with the pattern represented by the PWM. It reports only those hits that are present in equivalent positions of both sequences and exceed a specified threshold score in both, AND are found in regions of the alignment above the specified. <>= library(Biostrings) data(MA0003.2) pwm <- toPWM(MA0003.2) aln1 <- DNAString("ACTTCACCAGCTCCCTGGCGGTAAGTTGATC---AAAGG---AAACGCAAAGTTTTCAAG") aln2 <- DNAString("GTTTCACTACTTCCTTTCGGGTAAGTAAATATATAAATATATAAAAATATAATTTTCATC") sitePairSet <- searchAln(pwm, aln1, aln2, seqname1="seq1", seqname2="seq2", min.score="50%", cutoff=0.5, strand="*", type="any") ## generate gff style output head(writeGFF3(sitePairSet)) head(writeGFF2(sitePairSet)) ## search the Axt alignment library(CNEr) axtFilesHg19DanRer7 <- file.path(system.file("extdata", package="CNEr"), "hg19.danRer7.net.axt") axtHg19DanRer7 <- readAxt(axtFilesHg19DanRer7) sitePairSet <- searchAln(pwm, axtHg19DanRer7, min.score="80%", windowSize=51L, cutoff=0.7, strand="*", type="any", conservation=NULL, mc.cores=2) GRangesTFBS <- toGRangesList(sitePairSet, axtHg19DanRer7) GRangesTFBS$targetTFBS GRangesTFBS$queryTFBS @ \subsection{searchPairBSgenome} \Rfunction{searchPairBSgenome} is designed to do the genome-wise phylogenetic footprinting. Given two \Rclass{BSgenome}, a chain file for liftover from one genome to another, \Rfunction{searchPairBSgenome} identifies the putative transcription factor binding sites which are conserved in both genomes. <>= library(rtracklayer) library(JASPAR2014) library(BSgenome.Hsapiens.UCSC.hg19) library(BSgenome.Mmusculus.UCSC.mm10) pfm <- getMatrixByID(JASPAR2014, ID="MA0004.1") pwm <- toPWM(pfm) chain <- import.chain("Downloads/hg19ToMm10.over.chain") sitePairSet <- searchPairBSgenome(pwm, BSgenome.Hsapiens.UCSC.hg19, BSgenome.Mmusculus.UCSC.mm10, chr1="chr1", chr2="chr1", min.score="90%", strand="+", chain=chain) @ \section{Use \textit{de novo} motif discovery software} In this section, we will introduce wrapper functions for external motif discovery programs. So far, \software{MEME} is supported. \subsection{\software{MEME}} \Rfunction{runMEME} takes a \Rclass{DNAStringSet} or a set of \Rclass{characters} as input, and returns a \Rclass{MotifSet} object. <>= motifSet <- runMEME(file.path(system.file("extdata", package="TFBSTools"), "crp0.s"), binary="meme", arguments=list("-nmotifs"=3) ) ## Get the sites sequences and surrounding sequences sitesSeq(motifSet, type="all") ## Get the sites sequences only sitesSeq(motifSet, type="none") consensusMatrix(motifSet) @ \section{Session info} The following is the session info that generated this vignette: <>= sessionInfo() @ \newpage \bibliography{TFBSTools} \end{document}