Title: Straightforward finding of differential expressions through intensive randomization in transcriptomic studies
Authors: Dorota Desaulle - University Paris Descartes (France) [presenting]
Bernard Hainque - University Paris Descartes (France)
Celine Hoffmann - University Paris Descartes (France)
Pascal Bigey - University Paris Descartes (France)
Yves Rozenholc - University Paris Descartes (France)
Abstract: Transcriptomic data measure proportions of $p$ transcripts to identify differential expressions (DE) from $n$ samples under two or more experimental conditions, through e.g. a multiple testing procedure. Unfortunately, statistical analysis is sensitive to the unknown individual fraction of tissue reacting in the biological experiment. Multiplicative normalization adjusts for this intrinsic sample variability before any differential comparisons. Apart from inadequate methods based on library size or housekeeping genes, normalizations try to find a proper subset of invariants to estimate the scaling factor. Such strategies fail on simple counter-examples by adjusting expression variabilities across the conditions and/or the DE. In this context of high dimensional data ($n<<p$), under the assumption that the majority of analyzed expressions is invariant, we propose a new procedure for finding DE. It is straightforward as it does not rely on a previous good normalization. Instead, the findings are obtained by iterating the following steps: random selection of a small normalization subset regardless its quality, multiple testing for discovery of DE on the normalized expressions controlled by FDR. After the iterations, decreasing sorted rates of detection are compared to upper bound of the probabilities of choosing a (wrong) subset containing at least one DE when the number of true DE varies to obtain the discoveries. Our procedure globally controls the FDR.