Re-analysis of publicly available datasets

Re-analysis of publicly available datasets Lucia Peixoto, Davide Risso, Shane G. Poplawski, Mathieu, E. Wimmer, Terence P. Speed, Marcelo A. Wood and Ted Abel We retrieved the pre-processed data of several publicly available studies from GEO (see main text for details). In this Section, we plot the PCA of each dataset using the original normalization. Starting from the data as normalized by the authors, or applying UQ scaling normalization if the authors provided only raw counts, we apply RUVs using all the genes as negative controls and choosing the value of k that led to the best looking RLE plot. For each dataset, we retained only the genes expressed in at least three replicate samples. This analysis is intended to show that published normalized datasets often show residual unwanted variation and that RUVs can remove unwanted variation when present and does not compromise the data when scaling normalization is working well. A more careful analysis of each dataset, e.g. by selecting a problem-specific set of negative control genes, could lead to better results.