This preview shows page 1. Sign up to view the full content.
Unformatted text preview: Cluster Analysis in DNA Microarray Experiments
Sandrine Dudoit and Robert Gentleman
Bioconductor Short Course Winter 2002 c Copyright 2002, all rights reserved
1 Outline Overview of cluster analysis. Clustering gene expression data. Clustering methods Partitioning methods. Hierarchical methods. Estimating the number of clusters. Other topics Inference. Outliers. Hybrid methods. Bagged clustering.
2 Supervised vs. unsupervised learning Task. Assign objects to classes on the basis of measurements made on these objects. Unsupervised learning. The classes are unknown a priori and need to be "discovered" from the data.
a.k.a. cluster analysis; class discovery; unsupervised pattern recognition. Supervised learning. The classes are predefined and the task is to understand the basis for the classification from a set of labeled objects. This information is then used to classify future observations.
a.k.a. classification; discriminant analysis; class prediction; supervised pattern recognition.
3 Cluster analysis Associated with each object is a set of G measurements which form the feature vector, X = (X1 , . . . , XG ). The feature vector X belongs to a feature space X (e.g., G ). The task is to identify groups, or clusters, of similar objects on the basis of a set of feature vectors, X1 = x1 , . . . , Xn = xn . 4 Cluster analysis Clustering is in some sense a more difficult problem than classification. In general, all the issues that must be addressed for classification must also be addressed for clustering. In addition, with clustering, there is no learning set of labeled observations; the number of groups is usually unknown; implicitly, one must have already selected both the relevant features and distance measure; the goals can be quite vague: "Find some interesting and important clusters in my data"; most of the algorithms that are appealing are computationally too complex to have exact solutions; approximate solutions are used instead and reproducibility becomes an issue.
5 Cluster analysis Clustering involves several distinct steps. First, a suitable distance between objects must be defined, based on relevant features. Then, a clustering algorithm must be selected and applied. The results of a clustering procedure can include both the number of clusters K (if not prespecified) and a set of n cluster labels {1, . . . , K} for the n objects to be clustered. Appropriate choices will depend on the questions being asked and available data.
6 Cluster analysis Clustering procedures fall into two broad categories. Hierarchical methods, either divisive or agglomerative. These methods provide a hierarchy of clusters, from the smallest, where all objects are in one cluster, through to the largest set, where each observation is in its own cluster. Partitioning methods. These usually require the specification of the number of clusters. Then, a mechanism for apportioning objects to clusters must be determined. Most methods used in practice are agglomerative hierarchical methods. In large part, this is due to the availability of efficient exact algorithms. 7 Distance The feature data are often transformed to an n n distance or similarity matrix, D = (dij ), for the n objects to be clustered. Once a distance measure between individual observations has been chosen, one must often also define a distance measure between clusters, or groups of observations (e.g., average, single, and complete linkage agglomeration). Different choices here can greatly affect the outcome. More details in the lecture Distances and Expression Measures. 8 R clustering software class package: Self Organizing Maps (SOM). cluster package: AGglomerative NESting (agnes), Clustering LARe Applications (clara), DIvisive ANAlysis (diana), Fuzzy Analysis (fanny), MONothetic Analysis (mona), Partitioning Around Medoids (pam). e1071 package: Fuzzy Cmeans clustering (cmeans), Bagged clustering (bclust). mva package: Hierarchical clustering (hclust, cophenetic), kmeans (kmeans). Specialized summary, plot, and print methods for clustering results.
9 Gene expression data Gene expression data on G genes (features) for n mRNA samples (observations) mRNA samples x12 x22 . . . xG2 ... ... .. . x1n x2n . . . XGn x 11 x21 = . . . xG1 Genes . . . xGn xgi = expression measure for gene g in mRNA sample i. An array of conormalized arrays. 10 Gene expression data Features correspond to expression levels of different genes; possible classes include tumor types (e.g., ALL, AML), clinical outcomes (survival, nonsurvival), and are labeled by {1, 2, . . . , K}. Gene expression data on G genes (features) for n mRNA samples (observations) xi yi = xi1 , xi2 , . . . , xiG gene expression profile / feature vector for sample i = class for sample i, i = 1, . . . , n. Other covariates such as age, sex may also be important and could be included in the analysis. 11 Gene expression data Most efforts to date have involved clustering only the expression measures collected on a number of different genes and samples. However, there is likely to be a need for incorporating other data into the analysis, such as sample level covariates and biological metadata. For example, a common task is to determine whether or not gene expression data can reliably identify or classify different types of a disease. However, one might ask as well whether such data improve our ability to classify over already available sample level covariate data.
12 Clustering gene expression data One can cluster genes and/or samples (arrays). Clustering leads to readily interpretable figures. Clustering strengthens the signal when averages are taken within clusters of genes (Eisen et al., 1998). Clustering can be helpful for identifying gene expression patterns in time or space. Clustering is useful, perhaps essential, when seeking new subclasses of cell samples (tumors, etc). Clustering can be used for quality control: compare array/gene clustering results to experimental variables such as array batch, mRNA amplification method, lab, experimenter, etc.
13 Clustering gene expression data Cluster genes (rows) to identify groups of coregulated genes, e.g., using a large number of yeast experiments; to identify spatial or temporal expression patterns; to reduce redundancy (cf. feature selection) in prediction; to detect experimental artifacts; for display purposes. Transformations of the expression data matrix using linear modeling may be useful in this context: genes arrays = genes estimated effects.
Cf. Microarray Experimental Design and Analysis, Summer 2002.
14 Clustering gene expression data Cluster samples or arrays (columns) to identify new classes of biological samples, e.g., new tumor classes, new cell types; to detect experimental artifacts; for display purposes. Cluster both rows and columns at once. 15 Clustering gene expression data Clustering can be employed for quality control purposes. The clusters that obtain from clustering arrays/genes should be compared with different experimental conditions such as: batch or production order of the arrays; batch of reagents; mRNA amplification procedure; technician; plate origin of clones, etc. Any relationships observed here should be considered as a potentially serious source of bias.
16 Tumor classification using gene expression data A reliable and precise classification of tumors is essential for successful diagnosis and treatment of cancer. Current methods for classifying human malignancies rely on a variety of morphological, clinical, and molecular variables. In spite of recent progress, there are still uncertainties in diagnosis. Also, it is likely that the existing classes are heterogeneous and comprise diseases which are molecularly distinct and follow different clinical courses. 17 Tumor classification using gene expression data DNA microarrays may be used to characterize the molecular variations among tumors by monitoring gene expression profiles on a genomic scale. This may lead to a finer and more reliable classification of tumors, and to the identification of marker genes that distinguish among these classes. Eventual clinical implications include an improved ability to understand and predict cancer survival. 18 Tumor classification using gene expression data There are three main types of statistical problems associated with tumor classification: 1. the identification of new tumor classes using gene expression profiles unsupervised learning; 2. the classification of malignancies into known classes supervised learning; 3. the identification of marker genes that characterize the different tumor classes feature selection. 19 Example: Row and column clustering Figure 1: Alizadeh et al. (2000). Distinct types of diffuse large Bcell lymphoma identified by gene expression profiling. Nature.
20 Clustering gene expression data Preliminary questions Which genes / arrays to use? Which transformation/standardization? Which distance function? Which clustering algorithm? Answers will depend on the biological problem. 21 Clustering gene expression data Important questions (which are generic) How many clusters? How reliable are the clustering results? Statistical inference: distributional properties of clustering results. Assessing the strength/confidence of cluster assignments for individual observations; Assessing cluster homogeneity. 22 Partitioning methods Partition the data into a prespecified number K of mutually exclusive and exhaustive groups. Iteratively reallocate the observations to clusters until some criterion is met, e.g., minimize withincluster sumsofsquares. Examples: kmeans; extension to fuzzy kmeans; Partitioning Around Medoids PAM (Kaufman & Rousseeuw, 1990); SelfOrganizing Maps SOM (Kohonen, 2001); modelbased clustering, e.g., Gaussian mixtures in Fraley & Raftery (1998, 2000) and McLachlan et al. (2001).
23 Partitioning around medoids Partitioning around medoids or PAM of Kaufman and Rousseeuw (1990) is a partitioning method which operates on a distance matrix, e.g., Euclidean distance matrix. For a prespecified number of clusters K, the PAM procedure is based on the search for K representative objects, or medoids, among the observations to be clustered. After finding a set of K medoids, K clusters are constructed by assigning each observation to the nearest medoid. 24 Partitioning around medoids The goal is to find K medoids, M = (m1 , . . . , mK ), which minimize the sum of the distances of the observations to their closest medoid, that is, M = argminM
i min d(xi , mk ).
k PAM can be applied to general data types and tends to be more robust than kmeans. 25 Silhouette plots Rousseeuw (1987) suggested a graphical display, the silhouette plot, which can be used to: (i) select the number of clusters and (ii) assess how well individual observations are clustered. The silhouette width of observation i is defined as sili = (bi  ai )/ max(ai , bi ), where ai denotes the average distance between i and all other observations in the cluster to which i belongs, and bi denotes the minimum average distance of i to objects in other clusters. Intuitively, objects with large silhouette width sili are wellclustered, those with small sili tend to lie between clusters.
26 Silhouette plots For a given number of clusters K, the overall average silhouette width for the clustering is simply the average of sili over all observations i, sil = i sili /n. Kaufman & Rousseeuw suggest estimating the number of clusters K by that which gives the largest average silhouette width, sil. Note that silhouette widths may be computed for the results of any partitioning clustering algorithm. 27 Partitioning around medoids Silhouette plot for ALL AML data
0.20 0.00 0.05 0.10 0.15 15 24 20 9 3 4 1 26 14 19 17 7 18 31 30 33 32 28 12 25 22 Euclidean distance for scaled arrays, K=2, G=3,051 genes Figure 2: Golub et al. (1999) ALL AML data. Silhouette plot for PAM, red=ALL, blue=AML. 28 PAMSIL PAMSIL. van der Laan, Pollard, & Bryan (2001). Replace PAM criteria function with average silhouette. PAM Criteria Algorithm Starting values K Overall performance Splitting large clusters Outliers 
i PAMSIL
i mink d(xi , mk ) Build sili Steepest ascent Given or dataadaptive "Robust" Yes Ignore Steepest ascent PAM, random Given or dataadaptive "Efficient" No Identify 29 Hierarchical methods Hierarchical clustering methods produce a tree or dendrogram. They avoid specifying how many clusters are appropriate by providing a partition for each K. The partitions are obtained from cutting the tree at different levels. The tree can be built in two distinct ways bottomup: agglomerative clustering; topdown: divisive clustering. 30 Hierarchical methods Hierachical clustering dengrogram for ALL AML data
55 45 ALL 50 ALL Height AML ALL 40 AML ALL 35 ALL AML AML AML AML ALL ALL AML AML AML AML AML ALL ALL ALL 30 ALL ALL ALL ALL ALL ALL ALL ALL ALL ALL ALL 25 d Average linkage, Euclidean distance for scaled arrays, G=3,051 genes Figure 3: Golub et al. (1999) ALL AML data. Dendrogram for agglomerative hierarchical clustering. 31 ALL ALL ALL ALL ALL ALL Agglomerative methods Start with n mRNA sample (or G gene) clusters. At each step, merge the two closest clusters using a measure of betweencluster distance which reflects the shape of the clusters. Betweencluster distance measures: Average linkage: average of pairwise distances; Single linkage: minimum of pairwise distances; Complete linkage: maximum of pairwise distances. More details are given in the lecture Distances and Expression Measures. 32 Divisive methods Start with only one cluster. At each step, split clusters into two parts. Advantages: Obtain the main structure of the data, i.e., focus on upper levels of dendrogram. Disadvantages: Computational difficulties when considering all possible divisions into two groups. Examples SelfOrganizing Tree Algorithm SOTA (Dopazo & Carazo, 1997); DIvisive ANAlysis DIANA (Kaufman & Rousseeuw, 1990).
33 Dendrograms Dendrograms are often used to visualize the nested sequence of clusters resulting from hierarchical clustering. While dendrograms are quite appealing because of their apparent ease of interpretation, they can be misleading. First, the dendrogram corresponding to a given hierarchical clustering is not unique, since for each merge one needs to specify which subtree should go on the left and which on the right there are 2(n1) different dendrograms. The default in the R function hclust (cluster package) is to order the subtrees so that the tighter cluster is on the left.
34 Dendrograms A second, and perhaps less recognized shortcoming of dendrograms, is that they impose structure on the data, instead of revealing structure in these data. Such a representation will be valid only to the extent that the pairwise distances possess the hierarchical structure imposed by the clustering algorithm. 35 Dendrograms The cophenetic correlation coefficient can be used to measure how well the hierarchical structure from the dendrogram represents the actual distances. This measure is defined as the correlation between the n(n  1)/2 pairwise dissimilarities between observations and their cophenetic distances from the dendrogram, i.e., the between cluster dissimilarities at which two observations are first joined together. The cophenetic distances have a great deal of structure, e.g., there are many ties. Function cophenetic in mva package.
36 Partitioning vs. hierarchical Partitioning Advantages: Provides clusters that satisfy an optimality criterion (approximately). Disadvantages: Need initial K, long computation time. Hierarchical Advantages: Fast computation (for agglomerative clustering). Disadvantages: Rigid, cannot correct later for erroneous decisions made earlier. 37 Estimating the number of clusters Internal indices. Statistics based on within and betweenclusters matrices of sumsofsquares and crossproducts (30 methods reviewed in Milligan & Cooper (1985)). Estimate is the number of clusters K which minimizes or maximizes one of these indices. Average silhouette width. (Kaufman & Rousseeuw, 1990). Modelbased methods. EM algorithm for Gaussian mixtures, Fraley & Raftery (1998, 2000) and McLachlan et al. (2001). Gap statistic. (Tibshirani et al., 2001). Resampling method, for each K compare an observed internal index to its expected value under a reference distribution and look for K which maximizes the difference.
38 MSS Mean Silhouette Split MSS. (Pollard & van der Laan, 2002). Given K clusters, consider each cluster k = 1, . . . , K separately Apply the clustering algorithm to the elements of cluster k. Choose the number of child clusters that maximizes the average silhouette width. Call this maximum the split silhouette, SSk . Define the mean split silhouette as a measure of average cluster heterogeneity. K 1 SSk . M SS(K) = K
k=1 Choose the number of clusters K which minimizes M SS(K).
39 MSS Identifies finer structure in gene expression data. When clustering genes, existing criteria tend to identify global structure only. Provides a measure of cluster heterogeneity. Computationally easy. 40 Clest Clest. (Dudoit & Fridlyand, 2002). Resampling method which estimates the number of clusters based on prediction accuracy. For each number of clusters k, repeatedly randomly divide the original dataset into two nonoverlapping sets, a learning set Lb and a test set T b , b = 1, . . . , B. Apply the clustering algorithm to observations in the learning set Lb . Build a classifier using the class labels from the clustering. Apply the classifier to the test set T b . Compute a similarity score sk,b comparing the test set class labels from prediction and clustering. 41 Clest The similarity score for k clusters is the median of the B similarity scores: tk = median(sk,1 , , sk,B ). The number of clusters K is estimated by comparing the observed similarity score tk for each k to its expected value under a suitable reference distribution with K = 1. Applies to any partitioning algorithm and any classifier. Better suited for clustering samples than clustering genes. 42 Inference van der Laan & Bryan (2001). General framework for statistical inference in cluster analysis. View clustering as a deterministic rule that can be applied to parameters (or estimates thereof) of the distribution of gene expression measures. Parameters of interest include covariances between the expression measures of different genes. The parametric bootstrap can be used to study distributional properties (bias, variance) of the clustering results.
43 Outliers In classification it has often been found useful to define a class of outliers. This does not seem to have been extended to clustering. However, outlier detection is an important issue since outliers can greatly affect the betweencluster distances. Simple tests for outliers, such as identifying observations that are responsible for a disproportionate amount of the withincluster sumofsquares seems prudent. 44 Hybrid method HOPACH Hierarchical Ordered Partitioning And Collapsing Hybrid HOPACH (van der Laan & Pollard, 2001) Apply a partitioning algorithm iteratively to produce a hierarchical tree of clusters. At each node, a cluster is partitioned into two or more smaller clusters. Splits are not restricted to be binary. E.g., choose K based on average silhouette. 45 Hybrid method HOPACH Hierarchical. Can look at clusters at increasing levels of detail. Ordered. Ordering of the clusters and elements within clusters is dataadaptive and unique, performing better than other hierarchical algorithms. Clustering and ordering are based on the same distance function. The ordering of elements in any level can be used to reorder the data or distance matrices, and visualize the clustering structure. Partitioning. At each node, a cluster is split into two or more smaller clusters. Collapsing. Clusters can be collapsed at any level of the tree to join similar clusters and correct for errors made in the partitioning steps. Hybrid. Combines the strengths of both partitioning and hierarchical clustering methods.
46 Bagged clustering Leisch (1999). Hybrid method combining partitioning and hierarchical methods. A partitioning method is applied to bootstrap learning sets and the resulting partitions are combined by performing hierarchical clustering of the cluster centers. Dudoit & Fridlyand (2002). Apply a partitioning clustering method to bootstrap samples of the learning set. Combine the resulting partitions by (i) voting or (ii) the creation of a new distance matrix. Assess confidence in the clustering results using cluster votes. 47 Acknowledgments Brown Lab, Biochemistry, Stanford. Sabina Chiaretti, Dana Farber Cancer Institute. Jane Fridlyand, UCSF Cancer Center. Mark van der Laan, Biostatistics, UC Berkeley. Katie Pollard, Biostatistics,UC Berkeley. Yee Hwa (Jean) Yang, Statistics, UC Berkeley. 48 ...
View
Full
Document
This note was uploaded on 02/10/2012 for the course CSE 5615 taught by Professor Mitra during the Fall '11 term at FIT.
 Fall '11
 Mitra

Click to edit the document details