Hierarchical False Discovery Rates: Large-scale Inference for Plate-based High-throughput Phenotyping Methods by

Hierarchical False Discovery Rates: Large-scale Inference for Plate-based High-throughput Phenotyping Methods Hannes Bretschneider 2011 This thesis introduces Hierarchical false discovery rates, a new semi-parametric Bayesian method for the detection of causal links between a genotype and phenotype in high-throughput phenotypic studies. Hierarchical false discovery rates are designed for plate-based studies and incorporate a hierarchical Bayes structure that allows for heterogeneity in plate-wise means and variances, but simultaneously shares statistical strength between plates to aid estimation. The model also allows the combination of information from multiple replicates of the same genotype and estimates an empirical control, which functions as a baseline for the identiVcation of signiVcant cases. Hierarchical false discovery rates are formulated as a generative Bayesian model and do not require parametric speciVcations of the non-null density. Non-parametric density estimation is combined with a Gibbs sampler algorithm for inference. I use the Hierarchical false discovery rate to analyze a recent genome-wide study of alternative splicing of the gene let-2 in Caenorhabditis elegans and to identify a set of genes that are candidates for upor down-regulation of the ratio between the exon 9 and exon 10 isoforms of the let-2 gene.