Clustering and Classification In Gene Expression Data Carlo Colantuoni Slide Acknowledgements: Elizabeth Garrett-Mayer, Rafael Irizarry, Giovanni Parmigiani, David Madigan, Kevin Coombs, Richard Simon, Ingo Ruczinski.Classification based in part on Chapter 10 of Hand, Manilla, & Smyth and Chapter 7 of Han and Kamber

Data from Garber et al. PNAS (98), 2001.

Clustering is an exploratory tool to see who's running with who: Genes and Samples. “Unsupervized” NOT for classification of samples. NOT for identification of differentially expressed genes. Clustering

Clustering organizes things that are close into groups. What does it mean for two genes to be close? What does it mean for two samples to be close? Once we know this, how do we define groups? Hierarchical and K-Means Clustering

Distance We need a mathematical definition of distance between two points What are points? If each gene is a point, what is the mathematical definition of a point?

Points Gene1= (E 11, E 12, …, E 1N )’ Gene2= (E 21, E 22, …, E 2N )’ Sample1= (E 11, E 21, …, E G1 )’ Sample2= (E 12, E 22, …, E G2 )’ E gi =expression gene g, sample i G G N DATA MATRIX

Most Famous Distance Euclidean distance –Example distance between gene 1 and 2: –Sqrt of Sum of ( E 1i -E 2i ) 2, i=1,…,N When N is 2, this is distance as we know it: Baltimore DC Distance When N is 20,000 you have to think abstractly

Correlation can also be used to compute distance Pearson Correlation Spearman Correlation Uncentered Correlation Absolute Value of Correlation

The difference is that, if you have two vectors X and Y with identical shape, but which are offset relative to each other by a fixed value, they will have a standard Pearson correlation (centered correlation) of 1 but will not have an uncentered correlation of 1.

The similarity/distance matrices 1 2 ………………………………...G G G G G 1 2 ……….N DATA MATRIX GENE SIMILARITY MATRIX

The similarity/distance matrices 1 2 …………..N G G 12...N12...N 1 2 ……….N DATA MATRIX SAMPLE SIMILARITY MATRIX

Gene and Sample Selection Do you want all genes included? What to do about replicates from the same individual/tumor? Genes that contribute noise will affect your results. Including all genes: dendrogram can’t all be seen at the same time. Perhaps screen the genes?

Two commonly seen clustering approaches in gene expression data analysis Hierarchical clustering –Dendrogram (red-green picture) –Allows us to cluster both genes and samples in one picture and see whole dataset “organized” K-means/K-medoids –Partitioning method –Requires user to define K = # of clusters a priori –No picture to (over)interpret

Hierarchical Clustering The most overused statistical method in gene expression analysis Gives us pretty red-green picture with patterns But, pretty picture tends to be pretty unstable. Many different ways to perform hierarchical clustering Tend to be sensitive to small changes in the data Provided with clusters of every size: where to “cut” the dendrogram is user-determined

Choose clustering direction Agglomerative clustering (bottom-up) –Starts with as each gene in its own cluster –Joins the two most similar clusters –Then, joins next two most similar clusters –Continues until all genes are in one cluster Divisive clustering (top-down) –Starts with all genes in one cluster –Choose split so that genes in the two clusters are most similar (maximize “distance” between clusters) –Find next split in same manner –Continue until all genes are in single gene clusters

Choose linkage method (if bottom-up) Single Linkage: join clusters whose distance between closest genes is smallest (elliptical) Complete Linkage: join clusters whose distance between furthest genes is smallest (spherical) Average Linkage : join clusters whose average distance is the smallest.

Dendrogram Creation + Interpretation

Cluster Assignment

450 relevant genes “noise” genes.450 relevant genes. Simulated Data with 4 clusters: 1-10, 11-20, 21-30, 31-40

K-means and K-medoids Partitioning Method Don’t get pretty picture MUST choose number of clusters K a priori More of a “black box” because output is most commonly looked at purely as assignments Each object (gene or sample) gets assigned to a cluster Begin with initial partition Iterate so that objects within clusters are most similar

K-means (continued) Euclidean distance most often used Spherical clusters. Can be hard to choose or figure out K. Not unique solution: clustering can depend on initial partition No pretty figure to (over)interpret

K-means Algorithm 1. Choose K centroids at random 2. Make initial partition of objects into k clusters by assigning objects to closest centroid 3.Calculate the centroid (mean) of each of the k clusters. 4.a. For object i, calculate its distance to each of the centroids. b. Allocate object i to cluster with closest centroid. c. If object was reallocated, recalculate centroids based on new clusters. 4. Repeat 3 for object i = 1,….N. 5.Repeat 3 and 4 until no reallocations occur. 6.Assess cluster structure for fit and stability

K-means We start with some data Interpretation: –We are showing expression for two samples for 14 genes –We are showing expression for two genes for 14 samples This is with 2 genes. Iteration = 0

K-means Choose K centroids These are starting values that the user picks. There are some data driven ways to do it Iteration = 0

K-means Make first partition by finding the closest centroid for each point This is where distance is used Iteration = 1

K-means Now re-compute the centroids by taking the middle of each cluster Iteration = 2

K-means Repeat until the centroids stop moving or until you get tired of waiting Iteration = 3

K-means Limitations Final results depend on starting values How do we chose K? There are methods but not much theory saying what is best. Where are the pretty pictures?

Assessing cluster fit and stability Most often ignored. Cluster structure is treated as reliable and precise Can be VERY sensitive to noise and to outliers Homogeneity and Separation Cluster Silhouettes: how similar genes within a cluster are to genes in other clusters (Rousseeuw Journal of Computation and Applied Mathematics, 1987)

Silhouettes Silhouette of gene i is defined as: a i = average distance of gene i to other gene in same cluster b i = average distance of gene i to genes in its nearest neighbor cluster

WADP: Weighted Average Discrepancy Pairs Add perturbations to original data Calculate the number of paired samples that cluster together in the original cluster that didn’t in the perturbed Repeat for every cutoff (i.e. for each k) Do iteratively Estimate for each k the proportion of discrepant pairs.

Classification Diagnostic tests are good examples of classifiers –A patient has a given disease or not –The classifier is a machine that accepts some clinical parameters as input, and spits out an prediction for the patient D Not-D Classes must be mutually exclusive and exhaustive

Components of Class Prediction Select features (genes) –Which genes will be included in the model Select type of classifier –E.g. (D)LDA, SVM, k-Nearest-Neighbor, … Fit parameters for model (train the classifier) Quantify predictive accuracy: Cross- Validation

Feature Selection Goal is to identify a small subset of genes which together give accurate predictions. Methods will vary depending on nature of classification problem –Choose genes with significant t-statistics to distinguish between two simple classes e.g.

Classifier Selection In microarray classification, the number of features is (almost) always much greater than the number of samples. Overfitting is a distinct risk, and increases with more complicated methods.

How microarrays differ from the rest of the world Complex classification algorithms such as neural networks that perform better elsewhere don’t do as well as simpler methods for expression data. Comparative studies have shown that simpler methods work as well or better for microarray problems because the number of candidate predictors exceeds the number of samples by orders of magnitude. (Dudoit, Fridlyand and Speed JASA 2001)

Statistical Methods Appropriate for Class Comparison may not be Appropriate for Class Prediction Demonstrating statistical significance of prognostic factors is not the same as demonstrating predictive accuracy. Demonstrating goodness of fit of a model to the data used to develop it is not a demonstration of predictive accuracy. Most statistical methods were not developed for p>>n prediction problems

Linear discriminant analysis If there are K classes, simply draw lines (planes) to divide the space of expression profiles into K regions, one for each class. If profile X falls in region K, predict class K.

Nearest Neighbor Classification To classify a new observation X, measure the distance d(X,X i ) between X and every sample X i in training set Assign to X the class label of its “nearest neighbor” in the training set.

Random Forests Build several “random” decision trees and have them vote to determine final classification

Evaluating a classifier Want to estimate the error rate when classifier is used to predict class of a new observation The ideal approach is to get a set of new observations, with known class label and see how frequently the classifier makes the correct prediction. Performance on the training set is a poor approach, and will deflate the error estimate. Cross validation methods are used to get less biased estimates of error using only the training data.

Split-Sample Evaluation Training-set –Used to select features, select model type, determine parameters and cut-off thresholds Test-set –Withheld until a single model is fully specified using the training-set. –Fully specified model is applied to the expression profiles in the test-set to predict class labels. –Number of errors is counted

V-fold cross validation Divide data into V groups. Hold one group back, train the classifier on other V-1 groups, and use it to predict the last one. Rotate through all V points, holding each back. Error estimate is total error rate on all V test groups.

Leave-one-out Cross Validation Hold one data point back, train the classifier on other n-1 data points, and use it to predict the last one. Rotate through all n points, holding each back. Error estimate is total error rate on all n test values.

training set test set specimens log-expression ratios specimens log-expression ratios full data set Non-cross-validated Prediction Cross-validated Prediction (Leave-one-out method) 1. Prediction rule is built using full data set. 2. Rule is applied to each specimen for class prediction. 1. Full data set is divided into training and test sets (test set contains 1 specimen). 2. Prediction rule is built from scratch using the training set. 3. Rule is applied to the specimen in the test set for class prediction. 4. Process is repeated until each specimen has appeared once in the test set.

Which to use depends mostly on sample size If the sample is large enough, split into test and train groups. If sample is barely adequate for either testing or training, use leave one out In between consider V-fold. This method can give more accurate estimates than leave one out, but reduces the size of training set.

Beware Cross-validation of a model cannot occur after selecting the genes to be used in the model

Incomplete (incorrect) Cross- Validation Publications are using all the data to select genes and then cross-validating only the parameter estimation component of model development –Highly biased –Many published complex methods which make strong claims based on incorrect cross-validation. Frequently seen in complex feature set selection algorithms Some software encourages inappropriate cross-validation

Gene-Expression Profiles in Hereditary Breast Cancer Breast tumors studied: 7 BRCA1+ tumors 8 BRCA2+ tumors 7 sporadic tumors Log-ratios measurements of 3226 genes for each tumor after initial data filtering cDNA Microarrays Parallel Gene Expression Analysis RESEARCH QUESTION Can we distinguish BRCA1+ from BRCA1– cancers and BRCA2+ from BRCA2– cancers based solely on their gene expression profiles?

BRCA1

BRCA2