WGCNA算法研究笔记

WGCNA方法的实例分析
    文章的作者将该算法用R实现，正好数模过后我也在进行matlab向R的转换，这套算法就成为了我自学R的良好素材(TIOBE今年三月公布的编程语言排行榜中，R列居第24位，超过了SAS和Matlab，看来自己的选择似乎不错)。
    关于该实例的数据和分析说明可以在以下网页中找到 http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork/Rpackages/WGCNA/ 由于我不想写成一篇翻译稿，因此只会点到代码中的一些关键环节，并解释每一步分析的作用。
    WGCNA所给的样本数据是F2代135只小鼠liver细胞的芯片杂交结果，需要以此进行module identification，并将建立好的module与性状关联。
    （1）数据的预处理：
    # Display the current working directory
    getwd();
    # If necessary, change the path below to the directory where the data files are stored.
    # 指定到数据所在目录
    workingDir = ".";
    setwd(workingDir);
    # Load the package
    # 这里需要保证计算机的R中安装了WGCNA这一分析包
    library(WGCNA);
    # The following setting is important, do not omit.
    options(stringsAsFactors = FALSE);
    #Read in the female liver data set
    femData = read.csv("LiverFemale3600.csv");
    # Take a quick look at what is in the data set:
    dim(femData);
    names(femData);
    # 从样本数据中抽取表达量数据
    # 有一点值得注意的是，样本数据中芯片表达量存在负值，为此我纠结可很长时间
    # 事实上，芯片表达数据在给出时是相对一个参比而言的，即将目的基因与参比基因的表达量相除
    # 然后取其对数值，得到每一个基因的表达量，因此如果目的基因的表达量低于参比基因
    # 那么其数据值将出现负数
    datExpr0 = as.data.frame(t(femData[, -c(1:8)]));
    names(datExpr0) = femData$substanceBXH;
    rownames(datExpr0) = names(femData)[-c(1:8)];
    # 检测异常值，个人觉得没什么太大用处，所有的基因都通过了检验
    gsg = goodSamplesGenes(datExpr0, verbose = 3);
    gsg$allOK
    if (!gsg$allOK)
    {
    # Optionally, print the gene and sample names that were removed:
    if (sum(!gsg$goodGenes)>0)
    printFlush(paste("Removing genes:", paste(names(datExpr0)[!gsg$goodGenes], collapse = ", ")));
    if (sum(!gsg$goodSamples)>0)
    printFlush(paste("Removing samples:", paste(rownames(datExpr0)[!gsg$goodSamples], collapse = ", ")));
    # Remove the offending genes and samples from the data:
    datExpr0 = datExpr0[gsg$goodSamples, gsg$goodGenes]
    }
    调用聚类函数，依据不同小鼠芯片的表达情况，对小鼠进行聚类，并作图
    sampleTree = flashClust(dist(datExpr0), method = "average");
    # Plot the sample tree: Open a graphic output window of size 12 by 9 inches
    # The user should change the dimensions if the window is too large or too small.
    sizeGrWindow(12,9)
    #pdf(file = "Plots/sampleClustering.pdf", width = 12, height = 9);
    par(cex = 0.6);
    par(mar = c(0,4,2,0))
    plot(sampleTree, main = "Sample clustering to detect outliers", sub="", xlab="", cex.lab = 1.5,
    cex.axis = 1.5, cex.main = 2)
    # 插入聚类分割线
    # Plot a line to show the cut
    abline(h = 15, col = "red");

# Determine cluster under the line
    clust = cutreeStatic(sampleTree, cutHeight = 15, minSize = 10)
    table(clust)
    # clust 1 contains the samples we want to keep.
    keepSamples = (clust==1)
    datExpr = datExpr0[keepSamples, ]
    nGenes = ncol(datExpr)
    nSamples = nrow(datExpr)
    # 导入性状信息
    traitData = read.csv("ClinicalTraits.csv");
    dim(traitData)
    names(traitData)
    # remove columns that hold information we do not need.
    # 提取有用的性状信息
    allTraits = traitData[, -c(31, 16)];
    allTraits = allTraits[, c(2, 11:36) ];
    dim(allTraits)
    names(allTraits)
    # Form a data frame analogous to expression data that will hold the clinical traits.
    femaleSamples = rownames(datExpr);
    traitRows = match(femaleSamples, allTraits$Mice);
    datTraits = allTraits[traitRows, -1];
    rownames(datTraits) = allTraits[traitRows, 1];
    collectGarbage();
    # 将小鼠的性状量化值在聚类图中表示出来
    # Re-cluster samples
    sampleTree2 = flashClust(dist(datExpr), method = "average")
    # Convert traits to a color representation: white means low, red means high, grey means missing entry
    traitColors = numbers2colors(datTraits, signed = FALSE);
    # Plot the sample dendrogram and the colors underneath.
    plotDendroAndColors(sampleTree2, traitColors,
    groupLabels = names(datTraits),
    main = "Sample dendrogram and trait heatmap")
    # 保存处理好的表达信息和性状信息
    save(datExpr, datTraits, file = "FemaleLiver-01-dataInput.RData")
    下图的上半部分是将135只小鼠按照其基因表达的相似性进行聚类的结果，下半部分为性状的可视化，采用heatmap的形式，将135只老鼠的性状量化值表示出来，颜色越红，其数值越大。

（2）网络的建立：
    为了简化这里采用automatic的方法。
    # 首先导入（1）步骤预处理好的数据
    # Display the current working directory
    getwd();
    # If necessary, change the path below to the directory where the data files are stored.
    # "." means current directory. On Windows use a forward slash / instead of the usual \.
    workingDir = ".";
    setwd(workingDir);
    # Load the WGCNA package
    library(WGCNA)
    # The following setting is important, do not omit.
    options(stringsAsFactors = FALSE);
    # Load the data saved in the first part
    lnames = load(file = "FemaleLiver-01-dataInput.RData");
    #The variable lnames contains the names of loaded variables.
    lnames
    # 选择一系列的power参数进行网络构建
    # 根据topology overlap criteria选择最佳的参数进行进一步分析
    # Choose a set of soft-thresholding powers
    powers = c(c(1:10), seq(from = 12, to=20, by=2))
    # Call the network topology analysis function
    sft = pickSoftThreshold(datExpr, powerVector = powers, verbose = 5)
    # Plot the results:
    sizeGrWindow(9, 5)
    par(mfrow = c(1,2));
    cex1 = 0.9;
    # Scale-free topology fit index as a function of the soft-thresholding power
    plot(sft$fitIndices[,1], -sign(sft$fitIndices[,3])*sft$fitIndices[,2],
    xlab="Soft Threshold (power)",ylab="Scale Free Topology Model Fit,signed R^2",type="n",
    main = paste("Scale independence"));
    text(sft$fitIndices[,1], -sign(sft$fitIndices[,3])*sft$fitIndices[,2],
    labels=powers,cex=cex1,col="red");
    # this line corresponds to using an R^2 cut-off of h
    abline(h=0.90,col="red")
    # Mean connectivity as a function of the soft-thresholding power
    plot(sft$fitIndices[,1], sft$fitIndices[,5],
    xlab="Soft Threshold (power)",ylab="Mean Connectivity", type="n",
    main = paste("Mean connectivity"))
    text(sft$fitIndices[,1], sft$fitIndices[,5], labels=powers, cex=cex1,col="red")
    # 选择好参数后，建立网络
    net = blockwiseModules(datExpr, power = 6, minModuleSize = 30,
    reassignThreshold = 0, mergeCutHeight = 0.25,
    numericLabels = TRUE, pamRespectsDendro = FALSE,
    saveTOMs = TRUE,
    saveTOMFileBase = "femaleMouseTOM",
    verbose = 3)

从图中可以看出回归系数 值与基因的连接数间的trade-off，随着power值得增加， 值不断增加，表明网络越接近于scale-free network，但是基因的连接数减少使得module identification逐渐失去意义。因此该问题有待进一步研究，建立一优化模型不失为一个好的选择。
    # 显示每一个module的基因数
    table(net$colors)
    # 绘制出基因的module分类图
    # open a graphics window
    sizeGrWindow(12, 9)
    # Convert labels to colors for plotting
    mergedColors = labels2colors(net$colors)
    # Plot the dendrogram and the module colors underneath
    plotDendroAndColors(net$dendrograms[[1]], mergedColors[net$blockGenes[[1]]],
    "Module colors",
    dendroLabels = FALSE, hang = 0.03,
    addGuide = TRUE, guideHang = 0.05)

如图，对芯片中所有基因按照TOM矩阵值进行聚类的结果，module color中每一种颜色对应聚类树上的基因属于一个module。
    # 保存识别好的module信息
    moduleLabels = net$colors
    moduleColors = labels2colors(net$colors)
    MEs = net$MEs;
    geneTree = net$dendrograms[[1]];
    save(MEs, moduleLabels, moduleColors, geneTree,
    file = "FemaleLiver-02-networkConstruction-auto.RData")