We construct an atlas of a gene expression databank, to visualize similarity relationships between expression data sets. Such an atlas could be used as an interface to the databank, for users searching for relevant background data or data for their own in-silico analyses. The two main research problems in constructing an atlas are (1) to preprocess the data to make different sets commensurable, and (2) to visualize the data. In this work we use only very simple preprocessing to study its feasibility, and focus on the visualization. We compare several recently introduced methods in the task, and show that a method called curvilinear components analysis outperforms the newer ones in terms of trustworthiness of the projections. The visualizations reveal the main sources of variation in the data, namely the differences between data sets, different labs, and different measurement methods, which supports feasibility of the visualization method in the task. The other conclusion is that better methods are needed for making the data sets commensurable. <\p> Last modified: Mon May 16 10:01:34 EEST 2005