Showing papers by "Siva Sivaganesan published in 2010"

A semi-parametric Bayesian model for unsupervised differential co-expression analysis

Abstract: Differential co-expression analysis is an emerging strategy for characterizing disease related dysregulation of gene expression regulatory networks. Given pre-defined sets of biological samples, such analysis aims at identifying genes that are co-expressed in one, but not in the other set of samples. We developed a novel probabilistic framework for jointly uncovering contexts (i.e. groups of samples) with specific co-expression patterns, and groups of genes with different co-expression patterns across such contexts. In contrast to current clustering and bi-clustering procedures, the implicit similarity measure in this model used for grouping biological samples is based on the clustering structure of genes within each sample and not on traditional measures of gene expression level similarities. Within this framework, biological samples with widely discordant expression patterns can be placed in the same context as long as the co-clustering structure of genes is concordant within these samples. To the best of our knowledge, this is the first method to date for unsupervised differential co-expression analysis in this generality. When applied to the problem of identifying molecular subtypes of breast cancer, our method identified reproducible patterns of differential co-expression across several independent expression datasets. Sample groupings induced by these patterns were highly informative of the disease outcome. Expression patterns of differentially co-expressed genes provided new insights into the complex nature of the ERα regulatory network. We demonstrated that the use of the co-clustering structure as the similarity measure in the unsupervised analysis of sample gene expression profiles provides valuable information about expression regulatory networks.

On Objective Priors for Testing Hypotheses About Some Poisson Models

Abstract: In the absence of prior information, use of non-informative proper priors is often crucial for testing hypotheses, when using the Bayesian approach. In this context, the use of objective priors such as intrinsic priors and Zellner’s g-priors have gained much interest. In this paper, we consider the use of these priors for testing hypotheses about means and regression coefficients when observations come from Poisson distributions. We first derive an intrinsic prior for testing the equality of several Poisson means. We then focus on g-priors, giving a new motivation, based on shrinkage and minimal training sample arguments, for a mixture g-prior recommended by Liang, Paulo, Molina, Clyde and Berger (2008) for normal linear models. Using the same motivation, we propose a mixture g-prior for Poisson regression model. While the proposed g prior is similar to the one used by Wang and George (2007), it is also different in certain aspects. Specifically, we show that the Bayes factor derived from the proposed prior is consistent. We also provide examples using simulated and real data.

Prior Information Based Bayesian Infinite Mixture Model.

Abstract: Unsupervised learning methods have been tremendously successful in extracting knowledge from genomics data generated by high throughput experimental assays. However, analysis of each dataset in isolation without incorporating potentially informative prior knowledge is limiting the utility of such procedures. Here we present a novel probabilistic model and computational algorithm for semi-supervised learning from genomics data. The probabilistic model is an extension of the Bayesian semiparametric Gaussian Infinite Mixture Model (GIMM) and training of model parameters is performed using Markov Chain Monte Carl algorithm. The utility of the procedure in improving precision of cluster analysis by incorporating prior information is demonstrated in a simulation study and the analysis of the real world genomics data.

Objective Bayesian Testing of a Poisson Mean

Abstract: We consider testing hypotheses about a single Poisson mean. When prior information is not available, use of objective priors is of interest. We provide intrinsic priors based on the arithmetic intrinsic and fractional Bayes factors, and evaluate their characteristics.