Comparison of Clustering Methods for Investigation of Genome-Wide Methylation Array Data

Abstract

The use of genome-wide methylation arrays has proved very informative to investigate both clinical and biological questions in human epigenomics. The use of clustering methods either for exploration of these data or to compare to an a priori grouping, e.g., normal versus disease allows assessment of groupings of data without user bias. However no consensus on the methods to use for clustering of methylation array approaches has been reached. To determine the most appropriate clustering method for analysis of illumina array methylation data, a collection of data sets was simulated and used to compare clustering methods. Both hierarchical clustering and non-hierarchical clustering methods (k-means, k-medoids, and fuzzy clustering algorithms) were compared using a range of distance and linkage methods. As no single method consistently outperformed others across different simulations, we propose a method to capture the best clustering outcome based on an additional measure, the silhouette width. This approach produced a consistently higher cluster accuracy compared to using any one method in isolation.

Single-Cell Transcript Profiles Reveal Multilineage Priming in Early Progenitors Derived from Lgr5 + Intestinal Stem Cells

“…K -means clustering of the RNA data, using the Silhouette measure (Clifford et al, 2011) to identify the best k (Fig. S1B), revealed two distinct cell populations that were roughly equal in size (Fig.…”

“…K -means clustering was done in MATLAB using the squared Euclidean distance of normalized data (z-scores). To determine the optimal k , we applied every value from 2 to 20, assessed the average Silhouette value (Clifford et al, 2011) for each clustering result (Figure S1B), and selected k =2, which gave the largest mean Silhouette value. Differentially expressed genes were identified using a two-sided Wilcoxon-Mann-Whitney rank sum test implemented in the “coin” package in R. Differences between populations were determined by subtracting median Ct values (equivalent to log 2 expression levels).…”

Antiepileptic drugs and the fetal epigenome

“…Multiscale bootstrap values for each node determined by resampling of 1,000 replicates are shown. Canberra distance and Ward Linkage was determined to provide the greatest cluster separation based on mean silhouette width using the ClusterRank software (Clifford et al., 2011).…”

“…This enables the exploration of complex data without the need for a priori definition of groups that may be biased by experimenter expectations. The most appropriate algorithm for clustering was determined empirically using the ClusterRank software (Clifford et al., 2011). To identify statistically significant differentially methylated sites, normalized data were log 2 transformed and a t ‐test with a false discovery rate (FDR) multiple hypothesis correction was conducted to compare the mean scores between identified clusters.…”

Computational Methods in Epigenetics

“…Another study used both techniques to cluster transcription factors [101]. Clifford et al [102] compared hierarchical clustering to other clustering techniques (k-means, k-medoids, and fuzzy clustering) to determine the most appropriate one for analyzing Illumina methylation data. Since no significant difference was found between the methods, a combination was proposed; the final output will be given by the method that achieves the best results in each case.…”