Empirical Bayes gene detection in Microarray analysis

This project concentrates on developing a nonparametric Empirical Bayes (EB) method on predicting patients' survival times based on the gene expression profiles. Since the number of redundant variables is large in microarray data, the traditional model selection criteria, such as AIC, BIC etc., do not work properly. The nonparametric EB method is used to find a new threshold based on an estimated latent distribution. This not only reduces the dimensions of the model but also increases the predictive accuracy of the selected model. We estimate a mixture model of chi-squared distribution with censored data via EM algorithm. The mathematical derivation and computational implementation are presented. A diffuse large B-cell lymphoma (DLBCL) dataset is used as an illustrating example in this project. Evaluation on an independent test data and two simulations have been done to demonstrate the methodology. The nonparametric EB method is a potentially powerful tool for detecting the effect of individual genes on survival times and for finding the best possible model in the a large-scaled dataset