Dear all! Please note this announcement of the next meeting of the Highlands Local RSS group, which will take place in Aberdeen in 2 weeks. As always everybody is welcome! Claus (local group secretary) ------------------------------------------------------------------------ SPEAKER: Natalia Bochkina (Edinburgh) TITLE: Objective and subjective approaches to analysis of differential expression in microarray data (joint work with Alex Lewin and Sylvia Richardson) DATE: 30 October 2007 TIME: 4pm (Tea from 3:45) VENUE: Strathcona Lecture Theatre, Rowett Research Institute, Greenburn Road, AB21 9SB (See details at: http://www.rowett.ac.uk/institute/location.html The Strathcona lecture theatre can be found within Strathcona Hall. This is the building to your right (opposite from the Reid Library/Reception), when you come to Greenburn Road from the A96.) ABSTRACT: We consider a problem of comparing the means in two conditions (such as disease versus healthy samples) for a large number of variables simultaneously given a small number of replicates, with motivation coming from analysis of gene expression data. Bayesian modelling of such data allows to "borrow strength" across variables to stabilise variance estimates given a small number of replicates, by using an exchangeable hierarchical prior for variable-specific variances. With regard to the mean, with consider two types of priors: noninformative ("objective") and mixture priors. To define a mixture prior, we use prior knowledge that there are 3 groups of variables: with equal means (the null hypothesis), with the mean in condition 1 greater than the mean in condition 2 (overexpressed) and with the reverse (underexpressed). Thus, we use a three component mixture distribution to model the difference between the means, with components being an atom at zero and gamma distributions with support on positive and on negative semiline. We study sensitivity to the choice of prior on simulated data and compare models with different priors using posterior predictive checks. Since the mixture model can be sensitive to the choice of prior, we also consider a Bayesian model with noninformative prior for the means. To find variables with unequal means, we propose adaptive interval hypothesis testing where the interval depends on variability of each variable. Since we are testing a large number of hypotheses simultaneously, we also propose an estimator of the false discovery rate which allows to control the number of false positives. The adaptive interval hypothesis testing can be extended to compound hypotheses, i.e. those involving more than one parameter. These approaches will be illustrated on gene expression data sets produced by BAIR consortium (www.bair.org.uk). -- *********************************************************************************** Dr Claus-D. Mayer | http://www.bioss.ac.uk Biomathematics & Statistics Scotland | email: [log in to unmask] Rowett Research Institute | Telephone: +44 (0) 1224 716652 Aberdeen AB21 9SB, Scotland, UK. | Fax: +44 (0) 1224 715349 ***********************************************************************************