BACKGROUND: Non-specific feature selection is a dimension reduction procedure performed prior to cluster analysis of high dimensional molecular data. Not all measured features are expected to show biological variation, so only the most varying are selected for analysis. In DNA methylation studies, DNA methylation is measured as a proportion, bounded between 0 and 1, with variance a function of the mean. Filtering on standard deviation biases the selection of probes to those with mean values near 0.5. We explore the effect this has on clustering, and develop alternate filter methods that utilize a variance stabilizing transformation for Beta distributed data and do not share this bias. RESULTS: We compared results for 11 different non-specific filters on eight Infinium HumanMethylation data sets, selected to span a variety of biological conditions. We found that for data sets having a small fraction of samples showing abnormal methylation of a subset of normally unmethylated CpGs, a characteristic of the CpG island methylator phenotype in cancer, a novel filter statistic that utilized a variance-stabilizing transformation for Beta distributed data outperformed the common filter of using standard deviation of the DNA methylation proportion, or its log-transformed M-value, in its ability to detect the cancer subtype in a cluster analysis. However, the standard deviation filter always performed among the best for distinguishing subgroups of normal tissue. The novel filter and standard deviation filter tended to favour features in different genome contexts; for the same data set, the novel filter always selected more features from CpG island promoters and the standard deviation filter always selected more features from non-CpG island intergenic regions. Interestingly, despite selecting largely non-overlapping sets of features, the two filters did find sample subsets that overlapped for some real data sets. CONCLUSIONS: We found two different filter statistics that tended to prioritize features with different characteristics, each performed well for identifying clusters of cancer and non-cancer tissue, and identifying a cancer CpG island hypermethylation phenotype. Since cluster analysis is for discovery, we would suggest trying both filters on any new data sets, evaluating the overlap of features selected and clusters discovered.

BACKGROUND: Etiologic studies of cancer increasingly use molecular features such as gene expression, DNA methylation and sequence mutation to subclassify the cancer type. In large population-based studies, the tumor tissues available for study are archival specimens that provide variable amounts of amplifiable DNA for molecular analysis. As molecular features measured from small amounts of tumor DNA are inherently noisy, we propose a novel approach to improve statistical efficiency when comparing groups of samples. We illustrate the phenomenon using the MethyLight technology, applying our proposed analysis to compare MLH1 DNA methylation levels in males and females studied in the Colon Cancer Family Registry. RESULTS: We introduce two methods for computing empirical weights to model heteroscedasticity that is caused by sampling variable quantities of DNA for molecular analysis. In a simulation study, we show that using these weights in a linear regression model is more powerful for identifying differentially methylated loci than standard regression analysis. The increase in power depends on the underlying relationship between variation in outcome measure and input DNA quantity in the study samples. CONCLUSIONS: Tumor characteristics measured from small amounts of tumor DNA are inherently noisy. We propose a statistical analysis that accounts for the measurement error due to sampling variation of the molecular feature and show how it can improve the power to detect differential characteristics between patient groups.

BACKGROUND: DNA methylation, a molecular feature used to investigate tumor heterogeneity, can be measured on many genomic regions using the MethyLight technology. Due to the combination of the underlying biology of DNA methylation and the MethyLight technology, the measurements, while being generated on a continuous scale, have a large number of 0 values. This suggests that conventional clustering methodology may not perform well on this data. RESULTS: We compare performance of existing methodology (such as k-means) with two novel methods that explicitly allow for the preponderance of values at 0. We also consider how the ability to successfully cluster such data depends upon the number of informative genes for which methylation is measured and the correlation structure of the methylation values for those genes. We show that when data is collected for a sufficient number of genes, our models do improve clustering performance compared to methods, such as k-means, that do not explicitly respect the supposed biological realities of the situation. CONCLUSION: The performance of analysis methods depends upon how well the assumptions of those methods reflect the properties of the data being analyzed. Differing technologies will lead to data with differing properties, and should therefore be analyzed differently. Consequently, it is prudent to give thought to what the properties of the data are likely to be, and which analysis method might therefore be likely to best capture those properties.

MOTIVATION: Aberrant DNA methylation is common in cancer. DNA methylation profiles differ between tumor types and subtypes and provide a powerful diagnostic tool for identifying clusters of samples and/or genes. DNA methylation data obtained with the quantitative, highly sensitive MethyLight technology is not normally distributed; it frequently contains an excess of zeros. Established tools to analyze this type of data do not exist. Here, we evaluate a variety of methods for cluster analysis to determine which is most reliable. RESULTS: We introduce a Bernoulli-lognormal mixture model for clustering DNA methylation data obtained using MethyLight. We model the outcomes using a two-part distribution having discrete and continuous components. It is compared with standard cluster analysis approaches for continuous data and for discrete data. In a simulation study, we find that the two-part model has the lowest classification error rate for mixture outcome data compared with other approaches. The methods are illustrated using DNA methylation data from a study of lung cancer cell lines. Compared with competing hierarchical clustering methods, the mixture model approaches have the lowest cross-validation error for detecting lung cancer subtype (non-small versus small cell). The Bernoulli-lognormal mixture assigns observations to subgroups with the lowest uncertainty. AVAILABILITY: Software is available upon request from the authors. SUPPLEMENTARY INFORMATION: http://www-rcf.usc.edu/~kims/SupplementaryInfo.html